{ "cells": [ { "cell_type": "code", "execution_count": 1, "id": "9bb7e9ef", "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "\n", "aussie_df = pd.read_csv(\"au_podcasts.csv\")" ] }, { "cell_type": "code", "execution_count": 2, "id": "9f6d5a3d", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "(2206, 13)\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
idurllinktitlenewest_item_dateoldest_item_dateepisode_countexplicitpublisher_namepublisher_emailavg_duration_stotal_duration_estimate_stotal_duration_estimate_h
04550https://anchor.fm/s/a5ae2f4/podcast/rsshttps://anchor.fm/comments-and-musingsComments and Musings10/12/202007/22/2015160Francis Lynchfrancislynch.me@gmail.com1591254567.1
14770https://anchor.fm/s/343ab914/podcast/rsshttps://anchor.fm/genericliveshowThe Generic Live Show03/28/202107/17/20171070Dale Campbellpodcasts60+343ab914@anchor.fm3495373965103.9
\n", "
" ], "text/plain": [ " id url \\\n", "0 4550 https://anchor.fm/s/a5ae2f4/podcast/rss \n", "1 4770 https://anchor.fm/s/343ab914/podcast/rss \n", "\n", " link title \\\n", "0 https://anchor.fm/comments-and-musings Comments and Musings \n", "1 https://anchor.fm/genericliveshow The Generic Live Show \n", "\n", " newest_item_date oldest_item_date episode_count explicit publisher_name \\\n", "0 10/12/2020 07/22/2015 16 0 Francis Lynch \n", "1 03/28/2021 07/17/2017 107 0 Dale Campbell \n", "\n", " publisher_email avg_duration_s total_duration_estimate_s \\\n", "0 francislynch.me@gmail.com 1591 25456 \n", "1 podcasts60+343ab914@anchor.fm 3495 373965 \n", "\n", " total_duration_estimate_h \n", "0 7.1 \n", "1 103.9 " ] }, "execution_count": 2, "metadata": {}, "output_type": "execute_result" } ], "source": [ "print(aussie_df.shape)\n", "aussie_df.head(2)" ] }, { "cell_type": "code", "execution_count": 836, "id": "a8dfe69f", "metadata": {}, "outputs": [], "source": [ "# 20 podcasts, 5 episodes, 1 clips each\n", "\n", "sample_id_list = [\n", " 966750, # Bevo\n", " 965640,\n", " 970010,\n", " 1181554,\n", " 802920,\n", " \n", " 493314,\n", " 638196,\n", " 764574,\n", " 1139063,\n", " 1145814, \n", "\n", "]\n", "\n", "sample_df = aussie_df[\n", " aussie_df[\"id\"].isin(set(sample_id_list))\n", "]\n", "\n", "assert(sample_df.shape[0] == 10)\n", "assert((sample_df[\"episode_count\"] >= 5).all())\n", "assert((sample_df[\"publisher_email\"].str.contains(\"anchor.fm\")).sum() == 0)" ] }, { "cell_type": "code", "execution_count": 843, "id": "737495d2", "metadata": {}, "outputs": [], "source": [ "# episode_metas = []\n", "episode_metas = pd.read_csv(\"5_ep_aussie_100.csv\").to_dict(orient=\"records\")" ] }, { "cell_type": "code", "execution_count": 844, "id": "ac8cb88a", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "10it [00:01, 6.25it/s]\n" ] } ], "source": [ "import urllib\n", "import re\n", "import feedparser\n", "import requests\n", "import tqdm\n", "import os\n", "import bs4\n", "from suno_utils.utils.text import normalize_whitespace\n", "\n", "\n", "DEFAULT_HEADER = {\n", " \"User-Agent\": (\n", " \"Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) \"\n", " \"Chrome/50.0.2661.75 Safari/537.36\"\n", " ),\n", " \"X-Requested-With\": \"XMLHttpRequest\"\n", "}\n", "\n", "\n", "n_episodes = 1\n", "# episode_metas = []\n", "for _, row in tqdm.tqdm(sample_df.iterrows()):\n", " if row[\"id\"] in [e[\"podcast_id\"] for e in episode_metas]:\n", " continue\n", " r = requests.get(row[\"url\"], headers=DEFAULT_HEADER)\n", " feed = feedparser.parse(row[\"url\"])\n", " for entry in feed[\"entries\"][:n_episodes]:\n", " audio_links = [l[\"href\"] for l in entry[\"links\"] if l[\"rel\"] == \"enclosure\"]\n", " unquoted = urllib.parse.unquote(audio_links[0])\n", " episode_url = re.match(r\".*\\/play\\/.*?\\/(https?\\:\\/\\/.*)\", unquoted).group(1)\n", " to_filepath = \"tmp/\" + os.path.basename(episode_url)\n", " summary = normalize_whitespace(bs4.BeautifulSoup(entry[\"summary\"]).text)\n", " # get episode if needed\n", " if not os.path.exists(to_filepath):\n", " r = requests.get(episode_url, headers=DEFAULT_HEADER)\n", " with open(to_filepath, \"wb\") as f:\n", " f.write(r.content)\n", " episode_metas.append({\n", " \"podcast_id\": row[\"id\"], \n", " \"id\": entry[\"id\"], \n", " \"filepath\": to_filepath,\n", " \"author\": entry[\"author\"], \n", " \"title\": entry[\"title\"], \n", " \"summary\": summary,\n", " })" ] }, { "cell_type": "code", "execution_count": 845, "id": "699b9a13", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'filepath': 'tmp/2557ebe3-f32e-77bf-0195-3f5007f6e455.mp3',\n", " 'author': 'Mick & Liam',\n", " 'title': 'Ghostbusters',\n", " 'summary': \"It's a Halloween treat! In this ghostly episode, the boys discuss the most spooktacular movie franchise (and the importance of small business) while taking you through the gateway to another dimension of movie music. There is no Mick & Liam, only Zuul.\",\n", " 'file_id': '802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000',\n", " 'wav_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.opus',\n", " 'hotwords': \"['Mick & Liam', 'Zuul']\",\n", " 'order_nr': 'TC0136719736'}" ] }, "execution_count": 845, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[0]" ] }, { "cell_type": "code", "execution_count": 846, "id": "b578e0f0", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 1145814,\n", " 'id': '66f71ada-8e6a-17ea-04ca-86f176a4f68f',\n", " 'filepath': 'tmp/13464287-44100-2-bcc890dea5e74.mp3',\n", " 'author': 'Aaron Schultz',\n", " 'title': \"#7 - Chris O'Neill\",\n", " 'summary': 'On today’s episode I’m chatting with my Chasing Ghosts bandmate Chris O’Neil. When thinking of people to have as guests on the show Chris was definitely at the top of that list. He’s seriously one of the most creative and talented people I know. He’s one of those guys that can do everything… really well. An excellent pianist, guitarist, vocalist, songwriter and we get into all that in our conversation.. but what I was really interested in is his business FrostFire Audio and his studio setup in South Melbourne. This is where he records soundtracks, audio, voice-overs and sound effects for video games, television commercials, film etc.. and has worked with some of the world’s biggest brands like Google, Panasonic, Yamaha, Honda, BBC, Konami, UNIQLO and heaps more. And as I record this he is trekking through Nepal with a group of fundraisers to help build a school. So, not only is he a super talented guy, he’s another absolute legend of a human being. Links to everything can be found in the show notes at thewritersblock.co and don’t forget to check Chris out on Instagram at @chrisoneilffa or on his website at frostfireaudio.com.au'}" ] }, "execution_count": 846, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[-1]" ] }, { "cell_type": "code", "execution_count": 664, "id": "39ca1773", "metadata": {}, "outputs": [], "source": [ "# !rm -rf episode_starts/*" ] }, { "cell_type": "code", "execution_count": 847, "id": "8e7fc85f", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 10/10 [00:35<00:00, 3.58s/it]\n" ] } ], "source": [ "for episode_meta in tqdm.tqdm(episode_metas):\n", " if \"order_nr\" in episode_meta:\n", " continue\n", " episode_meta[\"file_id\"] = \"{}__{}__{}__{}\".format(\n", " episode_meta[\"podcast_id\"], episode_meta[\"id\"], 0, 5*60*1_000\n", " )\n", " to_wav_filepath = \"episode_starts/{}.wav\".format(episode_meta[\"file_id\"])\n", " to_opus_filepath = \"episode_starts/{}.opus\".format(episode_meta[\"file_id\"])\n", " episode_meta[\"wav_filepath\"] = to_wav_filepath\n", " episode_meta[\"opus_filepath\"] = to_opus_filepath\n", " if os.path.exists(to_wav_filepath) and os.path.exists(to_opus_filepath):\n", " continue\n", " audio = load_ffmpeg(episode_meta[\"filepath\"], sample_rate=44_100)\n", " audio_slice = audio.get_slice(to_s=5*60)\n", " audio_slice.write_wav(to_wav_filepath)\n", " write_opus(audio_slice, to_opus_filepath)" ] }, { "cell_type": "code", "execution_count": 848, "id": "b097a7a3", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'filepath': 'tmp/2557ebe3-f32e-77bf-0195-3f5007f6e455.mp3',\n", " 'author': 'Mick & Liam',\n", " 'title': 'Ghostbusters',\n", " 'summary': \"It's a Halloween treat! In this ghostly episode, the boys discuss the most spooktacular movie franchise (and the importance of small business) while taking you through the gateway to another dimension of movie music. There is no Mick & Liam, only Zuul.\",\n", " 'file_id': '802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000',\n", " 'wav_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.opus',\n", " 'hotwords': \"['Mick & Liam', 'Zuul']\",\n", " 'order_nr': 'TC0136719736'}" ] }, "execution_count": 848, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[0]" ] }, { "cell_type": "code", "execution_count": 849, "id": "62623bd5", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 1145814,\n", " 'id': '66f71ada-8e6a-17ea-04ca-86f176a4f68f',\n", " 'filepath': 'tmp/13464287-44100-2-bcc890dea5e74.mp3',\n", " 'author': 'Aaron Schultz',\n", " 'title': \"#7 - Chris O'Neill\",\n", " 'summary': 'On today’s episode I’m chatting with my Chasing Ghosts bandmate Chris O’Neil. When thinking of people to have as guests on the show Chris was definitely at the top of that list. He’s seriously one of the most creative and talented people I know. He’s one of those guys that can do everything… really well. An excellent pianist, guitarist, vocalist, songwriter and we get into all that in our conversation.. but what I was really interested in is his business FrostFire Audio and his studio setup in South Melbourne. This is where he records soundtracks, audio, voice-overs and sound effects for video games, television commercials, film etc.. and has worked with some of the world’s biggest brands like Google, Panasonic, Yamaha, Honda, BBC, Konami, UNIQLO and heaps more. And as I record this he is trekking through Nepal with a group of fundraisers to help build a school. So, not only is he a super talented guy, he’s another absolute legend of a human being. Links to everything can be found in the show notes at thewritersblock.co and don’t forget to check Chris out on Instagram at @chrisoneilffa or on his website at frostfireaudio.com.au',\n", " 'file_id': '1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000',\n", " 'wav_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.opus'}" ] }, "execution_count": 849, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[-1]" ] }, { "cell_type": "code", "execution_count": 667, "id": "68a20e54", "metadata": {}, "outputs": [], "source": [ "import spacy\n", "# from suno_utils.utils.text import make_unique_list\n", "\n", "nlp = spacy.load(\"en_core_web_lg\")\n", "\n", "def make_unique_list(input_list, key_f=None):\n", " el_seen = set()\n", " reduced_list = []\n", " for el in input_list:\n", " if key_f is None:\n", " k = el\n", " else:\n", " k = key_f(el)\n", " if k not in el_seen:\n", " reduced_list.append(el)\n", " el_seen.add(k)\n", " return reduced_list\n", "\n", "def get_hotwords_from_text(text):\n", " doc = nlp(text)\n", " hotwords = []\n", " for ent in doc.ents:\n", " if ent.label_ in {\"CARDINAL\", \"DATE\", \"MONEY\", \"ORDINAL\", \"PERCENT\", \"QUANTITY\", \"TIME\"}:\n", " continue\n", " hotwords.append(ent.text[:255])\n", " return hotwords" ] }, { "cell_type": "code", "execution_count": 850, "id": "3a48e379", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 10/10 [00:00<00:00, 84.96it/s]\n" ] } ], "source": [ "for episode_meta in tqdm.tqdm(episode_metas):\n", " if \"order_nr\" in episode_meta:\n", " continue\n", " hotwords = [episode_meta[\"author\"]]\n", " hotwords += get_hotwords_from_text(episode_meta[\"title\"] + \" \" + episode_meta[\"summary\"])\n", " hotwords = make_unique_list(hotwords)[:1000]\n", " episode_meta[\"hotwords\"] = hotwords" ] }, { "cell_type": "code", "execution_count": 704, "id": "74118692", "metadata": {}, "outputs": [], "source": [ "import json\n", "import time\n", "import random\n", "\n", "CLIENT_API_KEY = \"_CbCcpYaSu4DF8NuQJdWniWG2Ls\"\n", "USER_API_KEY = \"hYhwAFC3w39VN23lwpzn26Jy1CA=\"\n", "AUTH_STRING = f\"Rev {CLIENT_API_KEY}:{USER_API_KEY}\"\n", "\n", "def _upload_file(filepath, ref_str=None):\n", " url = \"https://api.rev.com/api/v1/inputs\"\n", " headers = {\n", " \"Authorization\": AUTH_STRING,\n", " \"Content-Type\": \"audio/ogg\",\n", " \"Content-Disposition\": \"attachment; filename=\\\"{}.{}\\\"\".format(ref_str, filepath.split(\".\")[-1]),\n", " }\n", " with open(filepath, 'rb') as f:\n", " data = f.read()\n", " r = requests.post(url, data=data, headers=headers)\n", " if r.status_code != 201:\n", " raise ValueError(r.reason)\n", " media_loc = r.headers[\"Location\"]\n", " return media_loc\n", "\n", "def _make_order(media_loc, ref_str=None, hotwords=None):\n", " headers = {\n", " \"Authorization\": AUTH_STRING,\n", " \"Content-Type\": \"application/json\",\n", " }\n", " url = \"https://api.rev.com/api/v1/orders\"\n", " hotwords = [] if hotwords is None else hotwords\n", " data = {\n", " \"sandbox_mode\": False,\n", " \"client_ref\": ref_str,\n", " \"transcription_options\": {\n", " \"inputs\": [\n", " {\n", " \"audio_length_seconds\": 60*5,\n", " \"uri\": media_loc,\n", " \"glossary\": hotwords,\n", " \"accents\": [\"Australian\"],\n", " },\n", " ],\n", " \"verbatim\": True,\n", " \"timestamps\": True,\n", " \"rush\": False,\n", " \"instant_first_draft\": False,\n", " \"output_file_formats\": [\"JSON\"],\n", " },\n", " }\n", "\n", " r = requests.post(url, data=json.dumps(data), headers=headers)\n", " if r.status_code != 201:\n", " raise ValueError(r.reason)\n", " order_number = r.headers[\"Location\"].split(\"/\")[-1]\n", " return order_number\n", "\n", "def _check_progress(order_number):\n", " headers = {\n", " \"Authorization\": AUTH_STRING,\n", " \"Content-Type\": \"application/json\",\n", " }\n", " url = \"https://api.rev.com/api/v1/orders/{}\".format(order_number)\n", " r = requests.get(url, headers=headers)\n", " if r.status_code != 200:\n", " raise ValueError(r.reason)\n", " if r.json()[\"status\"] != \"Complete\":\n", " return None\n", " else:\n", " transcript_id = [e for e in r.json()[\"attachments\"] if e[\"kind\"] == \"transcript\"][0][\"id\"]\n", " return transcript_id\n", " \n", "def _wait_till_done(order_number):\n", " while True:\n", " transcript_id = _check_progress(order_number)\n", " if transcript_id is not None:\n", " break\n", " time.sleep(60 * 5)\n", " return transcript_id\n", "\n", "def _get_transcript(transcript_id):\n", " headers = {\n", " \"Authorization\": AUTH_STRING,\n", " \"Accept\": \"application/json+rev-transcript\",\n", " }\n", " url = \"https://api.rev.com/api/v1/attachments/{}/content\".format(transcript_id)\n", " r = requests.get(url, headers=headers)\n", " if r.status_code != 200:\n", " raise ValueError(r.reason)\n", " return r.json()" ] }, { "cell_type": "code", "execution_count": 672, "id": "607c9e13", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'filepath': 'tmp/2557ebe3-f32e-77bf-0195-3f5007f6e455.mp3',\n", " 'author': 'Mick & Liam',\n", " 'title': 'Ghostbusters',\n", " 'summary': \"It's a Halloween treat! In this ghostly episode, the boys discuss the most spooktacular movie franchise (and the importance of small business) while taking you through the gateway to another dimension of movie music. There is no Mick & Liam, only Zuul.\",\n", " 'file_id': '802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000',\n", " 'wav_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.opus',\n", " 'hotwords': ['Mick & Liam', 'Zuul']}" ] }, "execution_count": 672, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[0]" ] }, { "cell_type": "code", "execution_count": 851, "id": "edf2cff2", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 1145814,\n", " 'id': '66f71ada-8e6a-17ea-04ca-86f176a4f68f',\n", " 'filepath': 'tmp/13464287-44100-2-bcc890dea5e74.mp3',\n", " 'author': 'Aaron Schultz',\n", " 'title': \"#7 - Chris O'Neill\",\n", " 'summary': 'On today’s episode I’m chatting with my Chasing Ghosts bandmate Chris O’Neil. When thinking of people to have as guests on the show Chris was definitely at the top of that list. He’s seriously one of the most creative and talented people I know. He’s one of those guys that can do everything… really well. An excellent pianist, guitarist, vocalist, songwriter and we get into all that in our conversation.. but what I was really interested in is his business FrostFire Audio and his studio setup in South Melbourne. This is where he records soundtracks, audio, voice-overs and sound effects for video games, television commercials, film etc.. and has worked with some of the world’s biggest brands like Google, Panasonic, Yamaha, Honda, BBC, Konami, UNIQLO and heaps more. And as I record this he is trekking through Nepal with a group of fundraisers to help build a school. So, not only is he a super talented guy, he’s another absolute legend of a human being. Links to everything can be found in the show notes at thewritersblock.co and don’t forget to check Chris out on Instagram at @chrisoneilffa or on his website at frostfireaudio.com.au',\n", " 'file_id': '1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000',\n", " 'wav_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.opus',\n", " 'hotwords': ['Aaron Schultz',\n", " 'Chris',\n", " 'FrostFire Audio',\n", " 'South Melbourne',\n", " 'Google',\n", " 'Panasonic',\n", " 'Yamaha',\n", " 'Honda',\n", " 'BBC',\n", " 'Konami',\n", " 'UNIQLO',\n", " 'Nepal',\n", " 'Instagram',\n", " '@chrisoneilffa',\n", " 'frostfireaudio.com.au']}" ] }, "execution_count": 851, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[-1]" ] }, { "cell_type": "code", "execution_count": 852, "id": "817fd4ef", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 10/10 [00:12<00:00, 1.25s/it]\n" ] } ], "source": [ "for episode_meta in tqdm.tqdm(episode_metas):\n", " if \"order_nr\" in episode_meta:\n", " continue\n", " media_loc = _upload_file(episode_meta[\"opus_filepath\"], ref_str=episode_meta[\"file_id\"])\n", " order_number = _make_order(media_loc, ref_str=episode_meta[\"file_id\"], hotwords=episode_meta[\"hotwords\"])\n", " episode_meta[\"order_nr\"] = order_number" ] }, { "cell_type": "code", "execution_count": 853, "id": "499403f5", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'filepath': 'tmp/2557ebe3-f32e-77bf-0195-3f5007f6e455.mp3',\n", " 'author': 'Mick & Liam',\n", " 'title': 'Ghostbusters',\n", " 'summary': \"It's a Halloween treat! In this ghostly episode, the boys discuss the most spooktacular movie franchise (and the importance of small business) while taking you through the gateway to another dimension of movie music. There is no Mick & Liam, only Zuul.\",\n", " 'file_id': '802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000',\n", " 'wav_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.opus',\n", " 'hotwords': \"['Mick & Liam', 'Zuul']\",\n", " 'order_nr': 'TC0136719736'}" ] }, "execution_count": 853, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[0]" ] }, { "cell_type": "code", "execution_count": 854, "id": "488acf82", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 1145814,\n", " 'id': '66f71ada-8e6a-17ea-04ca-86f176a4f68f',\n", " 'filepath': 'tmp/13464287-44100-2-bcc890dea5e74.mp3',\n", " 'author': 'Aaron Schultz',\n", " 'title': \"#7 - Chris O'Neill\",\n", " 'summary': 'On today’s episode I’m chatting with my Chasing Ghosts bandmate Chris O’Neil. When thinking of people to have as guests on the show Chris was definitely at the top of that list. He’s seriously one of the most creative and talented people I know. He’s one of those guys that can do everything… really well. An excellent pianist, guitarist, vocalist, songwriter and we get into all that in our conversation.. but what I was really interested in is his business FrostFire Audio and his studio setup in South Melbourne. This is where he records soundtracks, audio, voice-overs and sound effects for video games, television commercials, film etc.. and has worked with some of the world’s biggest brands like Google, Panasonic, Yamaha, Honda, BBC, Konami, UNIQLO and heaps more. And as I record this he is trekking through Nepal with a group of fundraisers to help build a school. So, not only is he a super talented guy, he’s another absolute legend of a human being. Links to everything can be found in the show notes at thewritersblock.co and don’t forget to check Chris out on Instagram at @chrisoneilffa or on his website at frostfireaudio.com.au',\n", " 'file_id': '1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000',\n", " 'wav_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/1145814__66f71ada-8e6a-17ea-04ca-86f176a4f68f__0__300000.opus',\n", " 'hotwords': ['Aaron Schultz',\n", " 'Chris',\n", " 'FrostFire Audio',\n", " 'South Melbourne',\n", " 'Google',\n", " 'Panasonic',\n", " 'Yamaha',\n", " 'Honda',\n", " 'BBC',\n", " 'Konami',\n", " 'UNIQLO',\n", " 'Nepal',\n", " 'Instagram',\n", " '@chrisoneilffa',\n", " 'frostfireaudio.com.au'],\n", " 'order_nr': 'TC0285203789'}" ] }, "execution_count": 854, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[-1]" ] }, { "cell_type": "code", "execution_count": 855, "id": "b4cf8fa0", "metadata": {}, "outputs": [], "source": [ "# pd.DataFrame(episode_metas).to_csv(\"10_ep_aussie_100.csv\", index=False)" ] }, { "cell_type": "code", "execution_count": 856, "id": "9e313477", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 10/10 [00:14<00:00, 1.49s/it]\n" ] } ], "source": [ "transcripts = []\n", "for episode_meta in tqdm.tqdm(episode_metas):\n", " order_number = episode_meta[\"order_nr\"]\n", " transcript_id = _wait_till_done(order_number)\n", " transcript_json = _get_transcript(transcript_id)\n", " transcripts.append(transcript_json)\n", " time.sleep(random.random() * 0.2)" ] }, { "cell_type": "code", "execution_count": 857, "id": "1f029b7e", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'filepath': 'tmp/2557ebe3-f32e-77bf-0195-3f5007f6e455.mp3',\n", " 'author': 'Mick & Liam',\n", " 'title': 'Ghostbusters',\n", " 'summary': \"It's a Halloween treat! In this ghostly episode, the boys discuss the most spooktacular movie franchise (and the importance of small business) while taking you through the gateway to another dimension of movie music. There is no Mick & Liam, only Zuul.\",\n", " 'file_id': '802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000',\n", " 'wav_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.wav',\n", " 'opus_filepath': 'episode_starts/802920__6f1afb09-65a6-488a-be33-e094f4677d86__0__300000.opus',\n", " 'hotwords': \"['Mick & Liam', 'Zuul']\",\n", " 'order_nr': 'TC0136719736'}" ] }, "execution_count": 857, "metadata": {}, "output_type": "execute_result" } ], "source": [ "episode_metas[0]" ] }, { "cell_type": "code", "execution_count": 901, "id": "d11b0da1", "metadata": {}, "outputs": [], "source": [ "def _parse_ts_as_s(timestamp):\n", " time_s = 0\n", " if \",\" in timestamp:\n", " time_s += float(\"0.\" + timestamp.split(\",\")[-1])\n", " timestamp = timestamp.split(\",\")[0]\n", " h, m, s = timestamp.split(\":\")\n", " time_s += int(s) + int(m) * 60 + int(h) * 60 * 60\n", " return round(time_s, 2)\n", "\n", "def _merge_tokens(tmp):\n", " assert(all([e[\"type\"] == \"text\" for e in tmp]))\n", " assert(len(set([e.get(\"speaker_id\") for e in tmp if e.get(\"speaker_id\") is not None])) == 1)\n", " merged_token = {\n", " \"type\": \"text\",\n", " \"value\": \"\".join([e[\"value\"] for e in tmp]),\n", " \"speaker_id\": tmp[0][\"speaker_id\"],\n", " }\n", " starts_s = [e.get(\"start_s\") for e in tmp if e.get(\"start_s\") is not None]\n", " ends_s = [e.get(\"end_s\") for e in tmp if e.get(\"end_s\") is not None]\n", " if len(starts_s) == 1 and len(ends_s) == 1:\n", " merged_token[\"start_s\"] = starts_s[0]\n", " merged_token[\"end_s\"] = ends_s[0]\n", " return merged_token\n", "\n", "def _parse_transcript(transcript):\n", " speaker_name_map = {e[\"id\"]: e[\"name\"] for e in transcript[\"speakers\"]}\n", " tokens = []\n", " for section in transcript[\"monologues\"]:\n", " speaker_id = section[\"speaker\"]\n", " for token in section[\"elements\"]:\n", " # normalize newlines/tabs\n", " token_text = re.sub(r\"\\s+\", \" \", token[\"value\"])\n", " if token[\"type\"] == \"tag\":\n", " tokens.append({\n", " \"type\": \"tag\",\n", " \"value\": token_text,\n", " \"include\": False,\n", " })\n", " continue\n", " elif token[\"type\"] != \"text\":\n", " raise ValueError(\"token type not understood\")\n", " if \" \" in token_text: \n", " for t in re.split(r\"( )\", token_text):\n", " if len(t) == 0:\n", " continue\n", " tokens.append({\n", " \"type\": \"text\",\n", " \"value\": t,\n", " \"speaker_id\": speaker_id,\n", " })\n", " continue\n", " if \"timestamp\" in token and \"end_timestamp\" in token:\n", " tokens.append({\n", " \"type\": \"text\",\n", " \"value\": token_text,\n", " \"speaker_id\": speaker_id,\n", " \"start_s\": _parse_ts_as_s(token[\"timestamp\"]),\n", " \"end_s\": _parse_ts_as_s(token[\"end_timestamp\"]),\n", " })\n", " else:\n", " tokens.append({\n", " \"type\": \"text\",\n", " \"value\": token_text,\n", " \"speaker_id\": speaker_id,\n", " })\n", " # end section with whitespace\n", " tokens.append({\n", " \"type\": \"text\",\n", " \"value\": \" \",\n", " \"speaker_id\": speaker_id,\n", " }) \n", " # do token merge logic\n", " merged_tokens = []\n", " tmp = []\n", " for token in tokens:\n", " if token[\"type\"] == \"tag\":\n", " if len(tmp) > 0:\n", " merged_tokens.append(_merge_tokens(tmp))\n", " tmp = []\n", " merged_tokens.append({\n", " \"type\": \"tag\",\n", " \"value\": token[\"value\"],\n", " \"include\": token[\"include\"],\n", " })\n", " elif token[\"type\"] == \"text\" and token[\"value\"] == \" \":\n", " if len(tmp) > 0:\n", " merged_tokens.append(_merge_tokens(tmp))\n", " tmp = []\n", " else:\n", " tmp.append(token.copy())\n", " if len(tmp) > 0:\n", " merged_tokens.append(_merge_tokens(tmp))\n", " tmp = []\n", " # add final touches like success and replacements\n", " out_tokens = []\n", " for token in merged_tokens:\n", " if \"start_s\" in token and \"end_s\" in token:\n", " token[\"success\"] = True\n", " else:\n", " token[\"success\"] = False\n", " # handle all \"tags\"\n", " if token[\"type\"] == \"tag\":\n", " out_tokens.append(token)\n", " continue\n", " elif token[\"type\"] != \"text\":\n", " raise ValueError()\n", " # TODO: what if for example \"(laughs)...\"\n", " elif \"(laughs)\" in token[\"value\"]:\n", " # handle trailing punctuation\n", " if token[\"value\"][-1:] == \".\" and len(out_tokens) >= 1:\n", " if out_tokens[-1][\"value\"][-1:] == \",\":\n", " out_tokens[-1][\"value\"] = out_tokens[-1][\"value\"][:-1] + \".\"\n", " elif out_tokens[-1][\"value\"][-1:] != \".\":\n", " out_tokens[-1][\"value\"] += \".\"\n", " # add tag token\n", " out_tokens.append({\n", " \"type\": \"tag\",\n", " \"value\": \"laughs\",\n", " \"success\": False,\n", " \"include\": True,\n", " })\n", " continue\n", " elif \"(singing)\" in token[\"value\"]:\n", " # handle trailing punctuation\n", " if token[\"value\"][-1:] == \".\" and len(out_tokens) >= 1:\n", " if out_tokens[-1][\"value\"][-1:] == \",\":\n", " out_tokens[-1][\"value\"] = out_tokens[-1][\"value\"][:-1] + \".\"\n", " elif out_tokens[-1][\"value\"][-1:] != \".\":\n", " out_tokens[-1][\"value\"] += \".\"\n", " # add tag token\n", " out_tokens.append({\n", " \"type\": \"tag\",\n", " \"value\": \"singing\",\n", " \"success\": False,\n", " \"include\": False,\n", " })\n", " continue\n", " elif \"(affirmative)\" in token[\"value\"]:\n", " # handle trailing punctuation\n", " if token[\"value\"][-1:] == \".\" and len(out_tokens) >= 1:\n", " if out_tokens[-1][\"value\"][-1:] == \",\":\n", " out_tokens[-1][\"value\"] = out_tokens[-1][\"value\"][:-1] + \".\"\n", " elif out_tokens[-1][\"value\"][-1:] != \".\":\n", " out_tokens[-1][\"value\"] += \".\"\n", " continue\n", " # replace hesitations\n", " token_text = token[\"value\"].replace(\"...\", \" -- \")\n", " token_text = re.sub(r\"\\-$\", \" --\", token_text)\n", " token_text = re.sub(r\"([Ww]h)\\-([Ww]h)\", \"\\\\1 -- \\\\2\", token_text)\n", " token_text = re.sub(r\"I\\-I\", \"I -- I\", token_text)\n", " token_text = re.sub(r\"([Tt])\\-([Tt])\", \"\\\\1 -- \\\\2\", token_text)\n", " # normalize some whitespace\n", " token_text = normalize_whitespace(token_text)\n", " # split off by whitespace again if needed\n", " if \" \" in token_text:\n", " for t in re.split(r\"( )\", token_text):\n", " if len(t) == 0:\n", " continue\n", " sub_token = {\n", " \"type\": \"text\",\n", " \"value\": t,\n", " \"speaker_id\": token[\"speaker_id\"],\n", " \"success\": False,\n", " }\n", " out_tokens.append(sub_token)\n", " else:\n", " out_tokens.append(token)\n", " # deal with tags\n", " # TODO: this is temporary and we will swap out with spans later\n", " final_tokens = []\n", " for token in out_tokens:\n", " if token[\"type\"] != \"tag\":\n", " final_tokens.append(token)\n", " continue\n", " if token[\"include\"]:\n", " continue\n", " final_tokens.append({\n", " \"type\": \"text\",\n", " \"value\": \"[meta]\",\n", " \"speaker_id\": -1,\n", " \"success\": False,\n", " })\n", " return speaker_name_map, final_tokens" ] }, { "cell_type": "code", "execution_count": 902, "id": "23ccbde6", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "85.3% tokens aligned\n" ] } ], "source": [ "speaker_names, tokens = _parse_transcript(transcripts[0])\n", "print(\"{}% tokens aligned\".format(round(np.mean([t[\"success\"] for t in tokens if t[\"type\"] == \"text\"]) * 100, 1)))" ] }, { "cell_type": "code", "execution_count": 903, "id": "f1e1fb27", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
typevaluespeaker_idstart_send_ssuccess
0textOkay.00.400.81True
1textThe01.421.50True
2textusual01.501.75True
3textstuff01.751.95True
4textisn't01.952.13True
\n", "
" ], "text/plain": [ " type value speaker_id start_s end_s success\n", "0 text Okay. 0 0.40 0.81 True\n", "1 text The 0 1.42 1.50 True\n", "2 text usual 0 1.50 1.75 True\n", "3 text stuff 0 1.75 1.95 True\n", "4 text isn't 0 1.95 2.13 True" ] }, "execution_count": 903, "metadata": {}, "output_type": "execute_result" } ], "source": [ "pd.DataFrame(tokens[:5])" ] }, { "cell_type": "code", "execution_count": null, "id": "39bd7408", "metadata": {}, "outputs": [], "source": [ "# TODO: \"film -- ... uh\" whitespace issues and still ... remaining" ] }, { "cell_type": "code", "execution_count": 834, "id": "bea16dc3", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[{'type': 'text',\n", " 'value': '...',\n", " 'speaker_id': 2,\n", " 'start_s': 47.0,\n", " 'end_s': 47.0,\n", " 'success': True},\n", " {'type': 'text',\n", " 'value': '...',\n", " 'speaker_id': 2,\n", " 'start_s': 48.6,\n", " 'end_s': 48.6,\n", " 'success': True}]" ] }, "execution_count": 834, "metadata": {}, "output_type": "execute_result" } ], "source": [ "[t for t in tokens if \"...\" in t[\"value\"]][:2]" ] }, { "cell_type": "code", "execution_count": null, "id": "3462e7a9", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "7deef768", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "85b6a794", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "c222b398", "metadata": {}, "source": [ "## Make slices" ] }, { "cell_type": "code", "execution_count": 904, "id": "e0e69b57", "metadata": {}, "outputs": [], "source": [ "from suno_utils.utils.slicer import get_align_gaps, get_intervals_from_gaps, get_vad_gaps" ] }, { "cell_type": "code", "execution_count": 905, "id": "e9884399", "metadata": {}, "outputs": [], "source": [ "def _get_slice_metas_tmp(audio, transcript, episode_meta):\n", " speaker_names, align_meta = _parse_transcript(transcript)\n", " # TODO: functionality to block a span, not crossable\n", " # TODO: change to start_s, end_s\n", " for e in align_meta:\n", " if \"start_s\" in e:\n", " e[\"start\"] = e[\"start_s\"]\n", " e[\"end\"] = e[\"end_s\"]\n", " \n", " vad_gaps = get_vad_gaps(audio, min_gap_s=0.1, vad_level=3)\n", " align_gaps = get_align_gaps(align_meta, audio.duration_s, min_gap_s=0.1, safe_pad=2)\n", "\n", " intervals = get_intervals_from_gaps(vad_gaps, align_gaps, max_unac_vad_tol_s=np.nan)\n", " \n", " slice_meta = [] # [(start, end), turns]\n", " for start_s, end_s in intervals:\n", " # skip everything pre-1min\n", " if start_s < 60:\n", " continue\n", " idx_start = None\n", " idx_end = None\n", " for n, e in enumerate(align_meta):\n", " if \"start_s\" in e and e[\"start_s\"] >= start_s and idx_start is None:\n", " idx_start = n\n", " if \"end_s\" in e and e[\"end_s\"] <= end_s:\n", " idx_end = n + 1\n", " if idx_start is None or idx_end is None:\n", " slice_tokens = []\n", " else:\n", " slice_tokens = align_meta[idx_start:idx_end]\n", " # split into speaker turns\n", " turns = []\n", " cur_speaker = None\n", " tmp = []\n", " for token in slice_tokens:\n", " if token[\"type\"] != \"text\":\n", " continue\n", " if token[\"speaker_id\"] != cur_speaker and len(tmp) > 0:\n", " turns.append((cur_speaker, \" \".join(tmp)))\n", " tmp = []\n", " tmp.append(token[\"value\"])\n", " cur_speaker = token[\"speaker_id\"]\n", " if len(tmp) > 0:\n", " turns.append((cur_speaker, \" \".join(tmp)))\n", " slice_meta.append(((start_s, end_s), turns))\n", "\n", " slice_cov_frac = np.sum([to_s - from_s for (from_s, to_s), turns in slice_meta]) / audio.duration_s\n", " print(\"{}% slice coverage\".format(round(slice_cov_frac * 100, 1)))\n", " \n", " return slice_meta" ] }, { "cell_type": "code", "execution_count": 816, "id": "b6f552da", "metadata": {}, "outputs": [], "source": [ "# !mkdir slices_aussie_100" ] }, { "cell_type": "code", "execution_count": 906, "id": "2f2f4aba", "metadata": { "scrolled": true }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "1it [00:01, 1.62s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "59.9% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "2it [00:03, 1.63s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "78.8% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "3it [00:04, 1.65s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "67.0% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "4it [00:06, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "54.4% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "5it [00:08, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "65.1% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "6it [00:09, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "69.7% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "7it [00:11, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "66.0% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "8it [00:13, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "68.5% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\r", "9it [00:14, 1.65s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "79.5% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "10it [00:16, 1.64s/it]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "67.5% slice coverage\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "full_slice_meta = []\n", "for transcript, episode_meta in tqdm.tqdm(zip(transcripts, episode_metas)):\n", " audio = Audio.from_file(episode_meta[\"wav_filepath\"])\n", " slice_meta = _get_slice_metas_tmp(audio, transcript, episode_meta)\n", " for (from_s, to_s), turns in slice_meta:\n", " # throw out if meta\n", " if \"[meta]\" in [e[-1] for e in turns]:\n", " continue\n", " clean_turns = []\n", " for sid, t in turns:\n", " clean_turns.append((sid, normalize_whitespace(t.replace(\"...\", \" -- \"))))\n", " slice_filepath = \"slices_aussie_100/{}__{}__{}__{}.wav\".format(\n", " episode_meta[\"podcast_id\"],\n", " episode_meta[\"id\"],\n", " int(round(from_s * 1_000)),\n", " int(round(to_s * 1_000)),\n", " )\n", " audio.get_slice(from_s=from_s, to_s=to_s).write_wav(slice_filepath)\n", " full_slice_meta.append({\n", " \"podcast_id\": episode_meta[\"podcast_id\"],\n", " \"id\": episode_meta[\"id\"],\n", " \"slice_filepath\": slice_filepath,\n", " \"start_s\": from_s,\n", " \"end_s\": to_s,\n", " \"turns\": clean_turns,\n", " })" ] }, { "cell_type": "code", "execution_count": 907, "id": "f44fd9d2", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "155 slices\n" ] } ], "source": [ "print(len(full_slice_meta), \"slices\")" ] }, { "cell_type": "code", "execution_count": 908, "id": "46107630", "metadata": {}, "outputs": [], "source": [ "with open(\"aussie_100_full_meta.json\", \"w\") as f:\n", " json.dump(full_slice_meta, f)" ] }, { "cell_type": "code", "execution_count": null, "id": "a67a9aff", "metadata": {}, "outputs": [], "source": [ "# TODO: mess up: So C-H-R-I-S-O-N-E-I-L-F-F-A.\"" ] }, { "cell_type": "code", "execution_count": 911, "id": "96d20c1f", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'podcast_id': 802920,\n", " 'id': '6f1afb09-65a6-488a-be33-e094f4677d86',\n", " 'slice_filepath': 'slices_aussie_100/802920__6f1afb09-65a6-488a-be33-e094f4677d86__64240__68090.wav',\n", " 'start_s': 64.24,\n", " 'end_s': 68.09,\n", " 'turns': [(2, 'Hm, well, coming from the')]}" ] }, "execution_count": 911, "metadata": {}, "output_type": "execute_result" } ], "source": [ "full_slice_meta[0]" ] }, { "cell_type": "code", "execution_count": 928, "id": "7fdf1591", "metadata": {}, "outputs": [], "source": [ "import shutil\n", "\n", "def _get_hash(s):\n", " return abs(hash(str(s))) % (10 ** 8)" ] }, { "cell_type": "code", "execution_count": 926, "id": "6f93eefb", "metadata": {}, "outputs": [], "source": [ "# !mkdir kiwi_corpus_sample\n", "# !mkdir kiwi_corpus_sample/slices" ] }, { "cell_type": "code", "execution_count": 946, "id": "980f633b", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "100 slices\n" ] } ], "source": [ "delive_slice_meta = []\n", "\n", "random.seed(6006)\n", "idx_list = list(range(len(full_slice_meta)))\n", "random.shuffle(idx_list)\n", "\n", "for idx in idx_list:\n", " slice_meta = full_slice_meta[idx]\n", " plaintext = \" \".join([e[-1] for e in slice_meta])\n", " if re.search(r\"[A-Z]\\-[A-Z]\\-[A-Z]\\-\", plaintext):\n", " continue\n", " slice_id = _get_hash(slice_meta[\"slice_filepath\"].split(\".\")[0].split(\"/\")[-1])\n", " from_fp = slice_meta[\"slice_filepath\"]\n", " slice_fp = \"slices/{}.wav\".format(slice_id)\n", " to_fp = \"kiwi_corpus_sample/\" + slice_fp\n", " shutil.copy(from_fp, to_fp)\n", " delive_slice_meta.append({\n", " \"source_id\": _get_hash(slice_meta[\"podcast_id\"]),\n", " \"id\": slice_id,\n", " \"filepath\": slice_fp,\n", " \"duration_s\": round(slice_meta[\"end_s\"] - slice_meta[\"start_s\"], 2),\n", " \"transcript\": [\n", " {\n", " \"speaker_id\": _get_hash(slice_meta[\"podcast_id\"] * 2) + speaker_id,\n", " \"text\": text,\n", " } for speaker_id, text in slice_meta[\"turns\"]\n", " ],\n", " })\n", " if len(delive_slice_meta) == 100:\n", " break\n", " \n", "print(len(delive_slice_meta), \"slices\")" ] }, { "cell_type": "code", "execution_count": 947, "id": "77dc7378", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "{\n", " \"source_id\": 61846527,\n", " \"id\": 55166786,\n", " \"filepath\": \"slices/55166786.wav\",\n", " \"duration_s\": 10.53,\n", " \"transcript\": [\n", " {\n", " \"speaker_id\": 55451236,\n", " \"text\": \"Uh, little bit of Ghostbusters 2.\"\n", " },\n", " {\n", " \"speaker_id\": 55451237,\n", " \"text\": \"Yup.\"\n", " },\n", " {\n", " \"speaker_id\": 55451236,\n", " \"text\": \"We're gonna pretend that the other one didn't happen, the third one. Um.\"\n", " },\n", " {\n", " \"speaker_id\": 55451237,\n", " \"text\": \"Yes.\"\n", " },\n", " {\n", " \"speaker_id\": 55451236,\n", " \"text\": \"The re -- the reboot.\"\n", " },\n", " {\n", " \"speaker_id\": 55451237,\n", " \"text\": \"The reboot.\"\n", " },\n", " {\n", " \"speaker_id\": 55451236,\n", " \"text\": \"Although, the reboot\"\n", " }\n", " ]\n", "}\n" ] } ], "source": [ "print(json.dumps(delive_slice_meta[1], indent=2))" ] }, { "cell_type": "code", "execution_count": 948, "id": "f3cae4a4", "metadata": {}, "outputs": [], "source": [ "with open(\"kiwi_corpus_sample/meta.jsonl\", \"w\") as f:\n", " for m in delive_slice_meta:\n", " f.write(json.dumps(m) + \"\\n\")" ] }, { "cell_type": "code", "execution_count": 937, "id": "3dd23aa1", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "/home/georg/notebooks/datasets/podcasts\r\n" ] } ], "source": [ "!cd /home/georg/notebooks/datasets/podcasts\n", "!tar -czf kiwi_corpus_sample.tar.gz kiwi_corpus_sample" ] }, { "cell_type": "code", "execution_count": null, "id": "c21b6fd7", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "9e7b60bc", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "f1d413fa", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "db3176b8", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 910, "id": "f8344361", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "149\n", "[(10, \"and we're sitting in your studio in south Melbourne. You just showed me some work that you've been working on that we're not allowed to talk about.\"), (11, 'Mm-hmm. Not yet, no.'), (10, \"Not yet. Um, but I'm pumped because it looks and sounds pretty awesome.\"), (11, 'Thank you.'), (10, 'Um,')]\n" ] }, { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "0\n", "[(2, 'Hm, well, coming from the')]\n" ] }, { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "80\n", "[(0, \"Um, there's gotta be some self-accountability on that end. Now, the whole point of this today is to say, is to get back to that old analogy, I can lead you to the water or I can encourage you and tell you all the reasons for it, but I can't make you actually do it.\")]\n" ] }, { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "127\n", "[(0, \"as well? So, Ronan, let's talk a little bit about the hyper normalization. Where do we kick off with this?\"), (1, \"We probably kick off in the fact that everybody would agree that it's costing them far more\")]\n" ] }, { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "36\n", "[(0, 'persuaded you to want to have a go at this new role?'), (1, 'No absolutely Bevo. So yeah, Furnace Park Football Club, yeah. So anything I was on their committee for 10 years and uh, anything from being the treasurer to looking after sponsors, getting more sponsors on board.')]\n" ] }, { "data": { "text/html": [ "\n", " \n", " " ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "idx_list = list(range(len(full_slice_meta)))\n", "random.shuffle(idx_list)\n", "\n", "for idx in idx_list[:5]:\n", " print(idx)\n", " slice_meta = full_slice_meta[idx]\n", " print(slice_meta[\"turns\"])\n", " Audio.from_file(slice_meta[\"slice_filepath\"]).play()\n", " print()" ] }, { "cell_type": "code", "execution_count": null, "id": "1d2e1868", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "e90e8d0f", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "10a4c391", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "612e707f", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "d7c7f60c", "metadata": {}, "source": [ "\"doing today? Very well, thanks. Well, today is either yesterday, today, or tomorrow, right? There are no weekends, it's just one of those three days. (laughs). Yup. We've got de ja vu all over again. Yup. Absolutely. Absolutely. Well, I- I- guess, uh,\"" ] }, { "cell_type": "code", "execution_count": null, "id": "661aefdb", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "1d907dbd", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "1733b3e3", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "1c313f0c", "metadata": {}, "source": [ "## Playground" ] }, { "cell_type": "code", "execution_count": 461, "id": "cb5c7191", "metadata": {}, "outputs": [], "source": [ "# all_order_numbers = [\n", "# \"TC0660919239\", # 0 - no timestamps\n", "# \"TC0822547396\", # 1 - no timestamps\n", "# \"TC1055915261\", # 2\n", "# \"TC0917365569\", # 3 - no timestamps\n", "# \"TC1004908015\", # 4\n", "# ]# + order_numbers" ] }, { "cell_type": "code", "execution_count": 156, "id": "202f20e1", "metadata": {}, "outputs": [], "source": [ "import os\n", "import tempfile\n", "\n", "import shlex\n", "import subprocess\n", "\n", "\n", "DEAFAULT_LD_LIBRARY_PATH = \"/usr/local/lib\"\n", "DEAFAULT_OPUSENC_PATH = \"/usr/local/bin/opusenc\"\n", "\n", "\n", "def write_opus(\n", " audio, \n", " filepath, \n", " bitrate=32, \n", " ld_libary_path=DEAFAULT_LD_LIBRARY_PATH, \n", " opusenc_path=DEAFAULT_OPUSENC_PATH\n", "):\n", " if filepath.split(\".\")[-1] != \"opus\":\n", " raise ValueError(\"output filepath needs to have a .opus extension\")\n", " with tempfile.TemporaryDirectory() as temp_dir:\n", " from_filepath = os.path.join(temp_dir, \"in.wav\")\n", " audio.write_wav(from_filepath)\n", " opus_env = os.environ.copy()\n", " opus_env[\"LD_LIBRARY_PATH\"] = ld_libary_path\n", " opus_convert_cmd = f\"{opusenc_path} --bitrate {bitrate} --speech {from_filepath} {filepath}\"\n", " out = subprocess.run(shlex.split(opus_convert_cmd), env=opus_env)\n", " if out.returncode != 0:\n", " raise ValueError(\"opus conversion failed\")" ] }, { "cell_type": "code", "execution_count": 148, "id": "6963ba61", "metadata": {}, "outputs": [], "source": [ "audio = load_ffmpeg(\"tmp/7fed0bcfd2c9a637837613af7bb1106e.m4a\", sample_rate=44_100)" ] }, { "cell_type": "code", "execution_count": 157, "id": "6b73ac80", "metadata": {}, "outputs": [], "source": [ "write_opus(audio.get_slice(to_s=5*60), \"test.opus\")" ] }, { "cell_type": "markdown", "id": "39657b5b", "metadata": {}, "source": [ "## Make slices" ] }, { "cell_type": "code", "execution_count": 73, "id": "c358a3a2", "metadata": {}, "outputs": [], "source": [ "from suno_utils.utils.notebook import load_ffmpeg\n", "from suno_utils.utils.slicer import get_intervals_from_gaps, get_vad_gaps\n", "import portion\n", "import numpy as np\n", "\n", "def get_intervals_from_gaps_new(\n", " vad_gap_intervals,\n", " min_dur_s=2.0,\n", " max_dur_s=15.0,\n", "):\n", " safe_gap_intervals = vad_gap_intervals\n", " slice_intervals = []\n", " offs_s = safe_gap_intervals.lower\n", " while True:\n", " gaps_found = safe_gap_intervals & portion.closed(offs_s + min_dur_s, offs_s + max_dur_s)\n", " if gaps_found.empty:\n", " remaining_interval = (safe_gap_intervals & portion.closedopen(offs_s + min_dur_s, np.inf))\n", " if not remaining_interval.empty:\n", " offs_s = remaining_interval.lower\n", " continue\n", " break\n", " # add slice\n", " slice_intervals.append((round(offs_s, 2), round(gaps_found.upper, 2)))\n", " offs_s = gaps_found.upper\n", " return slice_intervals" ] }, { "cell_type": "code", "execution_count": 58, "id": "fdb1a482", "metadata": {}, "outputs": [], "source": [ "episode_filepaths = os.listdir(\"tmp\")\n", "assert(len(episode_filepaths) == 100)" ] }, { "cell_type": "code", "execution_count": null, "id": "845c5982", "metadata": {}, "outputs": [], "source": [ "local_paths = []\n", "for n, filepath in tqdm.tqdm(enumerate(episode_filepaths)):\n", " audio = load_ffmpeg(\"tmp/\" + filepath, sample_rate=44_100)\n", " audio_16khz = audio.convert(16_000, 2, 1)\n", " vad_gap_intervals = get_vad_gaps(audio_16khz, vad_level=3)\n", " slice_intervals = get_intervals_from_gaps_new(vad_gap_intervals)\n", " lower_s, upper_s = slice_intervals[:20][-1]\n", " to_filepath = \"labeler/sample_100/{}.wav\".format(n)\n", " local_paths.append(to_filepath)\n", " audio.get_slice(from_s=lower_s, to_s=upper_s).write_wav(to_filepath)" ] }, { "cell_type": "code", "execution_count": null, "id": "392b5fa0", "metadata": {}, "outputs": [], "source": [ "# TODO: find hard slices!!" ] }, { "cell_type": "code", "execution_count": null, "id": "356281e3", "metadata": {}, "outputs": [], "source": [ "labeler_dir = \"labeler/\"" ] }, { "cell_type": "code", "execution_count": 98, "id": "50d52107", "metadata": {}, "outputs": [], "source": [ "from suno_utils.utils.notebook import transcribe\n", "from nemo.collections.nlp.models import PunctuationCapitalizationModel\n", "\n", "punct_model = PunctuationCapitalizationModel.from_pretrained(\"punctuation_en_bert\")" ] }, { "cell_type": "code", "execution_count": 113, "id": "03fd2a5c", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 100/100 [00:38<00:00, 2.62it/s]\n" ] } ], "source": [ "data = []\n", "for filepath in tqdm.tqdm(local_paths):\n", " norm_transcript_guess = transcribe(filepath)\n", " data.append((filepath, norm_transcript_guess))" ] }, { "cell_type": "code", "execution_count": 114, "id": "4612ad0a", "metadata": {}, "outputs": [], "source": [ "df = pd.DataFrame(data, columns=[\"audio_url\", \"norm_transcript_guess\"])" ] }, { "cell_type": "code", "execution_count": 115, "id": "3ca5f9e9", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2022-04-07 21:44:38 punctuation_capitalization_model:1068] Using batch size 100 for inference\n", "[NeMo I 2022-04-07 21:44:38 punctuation_capitalization_infer_dataset:91] Max length: 64\n", "[NeMo I 2022-04-07 21:44:38 data_preprocessing:358] Some stats of the lengths of the sequences:\n", "[NeMo I 2022-04-07 21:44:38 data_preprocessing:360] Min: 15 | Max: 80 | Mean: 46.5 | Median: 48.0\n", "[NeMo I 2022-04-07 21:44:38 data_preprocessing:366] 75 percentile: 55.00\n", "[NeMo I 2022-04-07 21:44:38 data_preprocessing:367] 99 percentile: 71.09\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|██████████| 2/2 [00:00<00:00, 5.36batch/s]\n" ] } ], "source": [ "df[\"orig_transcript_guess\"] = punct_model.add_punctuation_capitalization(df[\"norm_transcript_guess\"].tolist())" ] }, { "cell_type": "code", "execution_count": 116, "id": "62205b72", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
audio_urlnorm_transcript_guessorig_transcript_guess
0sample_100/0.wavbut money in itself does not make us happy lik...But money in itself does not make us happy. Li...
1sample_100/1.wavbut she hears allse things about her um okay s...But she hears allse things about her. Um, Okay...
\n", "
" ], "text/plain": [ " audio_url norm_transcript_guess \\\n", "0 sample_100/0.wav but money in itself does not make us happy lik... \n", "1 sample_100/1.wav but she hears allse things about her um okay s... \n", "\n", " orig_transcript_guess \n", "0 But money in itself does not make us happy. Li... \n", "1 But she hears allse things about her. Um, Okay... " ] }, "execution_count": 116, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df.head(2)" ] }, { "cell_type": "code", "execution_count": 118, "id": "820b2a86", "metadata": {}, "outputs": [], "source": [ "df.to_csv(\"labeler/sample_100.csv\", index=False)" ] }, { "cell_type": "code", "execution_count": null, "id": "c0c26c4a", "metadata": {}, "outputs": [], "source": [ "# TODO: add this to repo" ] }, { "cell_type": "code", "execution_count": null, "id": "9164379b", "metadata": {}, "outputs": [], "source": [ "# TODO: sample 8 ends mid-word" ] }, { "cell_type": "code", "execution_count": null, "id": "4881f822", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "b6b7ffee", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.8.10" } }, "nbformat": 4, "nbformat_minor": 5 }