{ "cells": [ { "cell_type": "code", "execution_count": 16, "id": "c8df4df8", "metadata": {}, "outputs": [], "source": [ "# try nemo cap/punct\n", "# try nemo ner\n", "# try spacy ner?\n", "# try spgi punct asr?\n", "# try whisper" ] }, { "cell_type": "code", "execution_count": 17, "id": "93d4a7a4", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "[NeMo W 2022-10-29 10:16:50 optimizers:67] Could not import distributed_fused_adam optimizer from Apex\n", "[NeMo W 2022-10-29 10:16:52 experimental:27] Module is experimental, not ready for production and is not fully supported. Use at your own risk.\n", "[NeMo W 2022-10-29 10:16:52 experimental:27] Module is experimental, not ready for production and is not fully supported. Use at your own risk.\n" ] } ], "source": [ "from IPython.utils.io import capture_output\n", "import re\n", "\n", "import nemo.collections.nlp as nemo_nlp\n", "import spacy\n", "import whisper\n", "import tqdm\n", "\n", "from suno_utils.utils.text import normalize_whitespace\n", "from suno_utils.audio import Audio\n", "from suno_utils.utils.display import suppress_logging" ] }, { "cell_type": "code", "execution_count": 18, "id": "182059d4", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "100% [......................................................................] 407269414 / 407269414" ] } ], "source": [ "with suppress_logging():\n", " punct_model = nemo_nlp.models.PunctuationCapitalizationModel.from_pretrained(\n", " model_name=\"punctuation_en_bert\",\n", " )\n", " punct_2_model = nemo_nlp.models.PunctuationCapitalizationModel.from_pretrained(\n", " model_name=\"punctuation_en_distilbert\",\n", " )\n", " ner_model = nemo_nlp.models.TokenClassificationModel.from_pretrained(\n", " model_name=\"ner_en_bert\",\n", " )\n", "def punct_predict(fp):\n", " with suppress_logging():\n", " with capture_output():\n", " punct_out = punct_model.add_punctuation_capitalization([fp])[0]\n", " return punct_out\n", "def punct_2_predict(fp):\n", " with suppress_logging():\n", " with capture_output():\n", " punct_out = punct_2_model.add_punctuation_capitalization([fp])[0]\n", " return punct_out\n", "NER_LABEL_MAP = {v: k for k, v in ner_model._cfg.label_ids.items()}\n", "def ner_predict(text):\n", " with suppress_logging():\n", " with capture_output():\n", " class_id_preds = ner_model._infer([text], 1)\n", " class_preds = [NER_LABEL_MAP[idx][2:] for idx in class_id_preds]\n", " tokens = text.split()\n", " assert(len(tokens) == len(class_preds))\n", " return list(zip(tokens, class_preds))" ] }, { "cell_type": "code", "execution_count": 19, "id": "e34d9f5c", "metadata": {}, "outputs": [], "source": [ "whisper_model = whisper.load_model(\"medium.en\")\n", "# whisper_2_model = whisper.load_model(\"large\")\n", "def whisper_predict(fp):\n", " audio = whisper.load_audio(fp)\n", " audio = whisper.pad_or_trim(audio)\n", " mels = whisper.log_mel_spectrogram(audio).to(whisper_model.device)\n", " options = whisper.DecodingOptions(language=\"en\", without_timestamps=True)\n", " result = whisper_model.decode(mels, options)\n", " return result.text\n", "# def whisper_2_predict(fp):\n", "# audio = whisper.load_audio(fp)\n", "# audio = whisper.pad_or_trim(audio)\n", "# mels = whisper.log_mel_spectrogram(audio).to(whisper_2_model.device)\n", "# options = whisper.DecodingOptions(language=\"en\", without_timestamps=True)\n", "# result = whisper_2_model.decode(mels, options)\n", "# return result.text" ] }, { "cell_type": "code", "execution_count": 20, "id": "ab55da94", "metadata": {}, "outputs": [], "source": [ "# nlp = spacy.load('en_core_web_lg')\n", "# def spacy_predict(text):\n", "# doc = nlp(text)\n", "# new_text = text\n", "# for ent in doc.ents[::-1]:\n", "# new_text = new_text[:ent.end_char] + f\"[{ent.label_}]\" + new_text[ent.end_char:]\n", "# return new_text" ] }, { "cell_type": "code", "execution_count": 21, "id": "7cbe5427", "metadata": {}, "outputs": [], "source": [ "data = []\n", "with open(\"transcripts.tsv\") as f:\n", " for line in f.read().strip().split(\"\\n\"):\n", " fp, s = line.split(\"\\t\")\n", " data.append((fp, s.lower()))\n", "text_list = [e[1] for e in data]\n", "fp_list = [e[0] for e in data]" ] }, { "cell_type": "code", "execution_count": 22, "id": "8e006d94", "metadata": {}, "outputs": [], "source": [ "punct_out = [punct_predict(text) for text in text_list]\n", "punct_2_out = [punct_2_predict(text) for text in text_list]" ] }, { "cell_type": "code", "execution_count": 23, "id": "5f38e17d", "metadata": {}, "outputs": [], "source": [ "ner_out = [ner_predict(text) for text in punct_out]" ] }, { "cell_type": "code", "execution_count": 24, "id": "3993c29c", "metadata": {}, "outputs": [], "source": [ "# ner_2_out = [spacy_predict(text) for text in text_list]\n", "# ner_2_post_out = [spacy_predict(text) for text in punct_out]" ] }, { "cell_type": "code", "execution_count": 25, "id": "83937286", "metadata": {}, "outputs": [], "source": [ "asr_out = [whisper_predict(fp) for fp in fp_list]\n", "# asr_2_out = [whisper_2_predict(fp) for fp in fp_list]" ] }, { "cell_type": "markdown", "id": "f3842095", "metadata": {}, "source": [ "## Consolidate" ] }, { "cell_type": "code", "execution_count": 26, "id": "87253ca6", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", "\n" ], "text/plain": [ "" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "Audio.from_file(fp_list[7]).play()" ] }, { "cell_type": "code", "execution_count": 27, "id": "d1afcf4b", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "['mustafa do you have a second',\n", " 'yo andy you got a second',\n", " 'seven eight nine',\n", " 'four five six',\n", " 'testing testing testing']" ] }, "execution_count": 27, "metadata": {}, "output_type": "execute_result" } ], "source": [ "text_list[:5]" ] }, { "cell_type": "code", "execution_count": 28, "id": "e80e5df0", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "['Mustafa, do you have a second?',\n", " 'Yo, Andy, you got a second?',\n", " 'Seven, eight nine.',\n", " 'Four, five, six.',\n", " 'Testing testing testing.']" ] }, "execution_count": 28, "metadata": {}, "output_type": "execute_result" } ], "source": [ "punct_out[:5]" ] }, { "cell_type": "code", "execution_count": 29, "id": "16324b94", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "['Mustafa, do you have a second?',\n", " 'Yo, Andy, you got a second.',\n", " 'Seven, eight, nine.',\n", " 'Four, five, six.',\n", " 'Testing testing testing.']" ] }, "execution_count": 29, "metadata": {}, "output_type": "execute_result" } ], "source": [ "punct_2_out[:5]" ] }, { "cell_type": "code", "execution_count": 30, "id": "120aef96", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[[('Mustafa,', 'PER'),\n", " ('do', ''),\n", " ('you', ''),\n", " ('have', ''),\n", " ('a', ''),\n", " ('second?', '')],\n", " [('Yo,', ''),\n", " ('Andy,', 'PER'),\n", " ('you', ''),\n", " ('got', ''),\n", " ('a', ''),\n", " ('second?', '')],\n", " [('Seven,', ''), ('eight', ''), ('nine.', 'TIME')],\n", " [('Four,', ''), ('five,', ''), ('six.', '')],\n", " [('Testing', ''), ('testing', ''), ('testing.', '')]]" ] }, "execution_count": 30, "metadata": {}, "output_type": "execute_result" } ], "source": [ "ner_out[:5]" ] }, { "cell_type": "code", "execution_count": 31, "id": "feba52a6", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "['Mustafar, do you have a second?',\n", " 'E-O-N-D, you got a second?',\n", " '789.',\n", " 'four five six.',\n", " 'Testing, testing, testing.']" ] }, "execution_count": 31, "metadata": {}, "output_type": "execute_result" } ], "source": [ "asr_out[:5]" ] }, { "cell_type": "code", "execution_count": 32, "id": "b7542147", "metadata": {}, "outputs": [], "source": [ "# use bert punctuator\n", "\n", "# punctuation\n", " # add distillbert commas\n", " # use whisper to fixup ambiguous puntuation and all-uppercase words like JSON and FBI\n", "\n", "# capitalization\n", " # find entities that remain uppercase\n", " # ??use nemo/spacy pre/post punct??\n", " # lowercase everything that is not entities\n", " # lowercase rules like numbers and 'i'" ] }, { "cell_type": "code", "execution_count": 34, "id": "44d826de", "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "50it [00:24, 2.04it/s]\n" ] } ], "source": [ "number_set = set([\n", " \"one\", \"two\", \"three\", \"four\", \"five\", \"six\", \"seven\", \"eight\", \"nine\", \"ten\", \n", " \"eleven\", \"twelve\", \"thirteen\", \"fourteen\", \"fifteen\", \"sixteen\", \"seventeen\", \"eighteen\", \"nineteen\"\n", "])\n", "\n", "def _lower(m):\n", " return \"<\" + m.group(1).lower() + \">\"\n", "\n", "out_texts = []\n", "for fp, text in tqdm.tqdm(zip(fp_list, text_list)):\n", " # TODO: handle background speech brackets (keep track of position)\n", " tokens = normalize_whitespace(text).split()\n", " # punctuate and capitalize\n", " with suppress_logging():\n", " with capture_output():\n", " punct_pred_1 = punct_model.add_punctuation_capitalization([text], return_labels=True)[0].split()\n", " punct_pred_2 = punct_2_model.add_punctuation_capitalization([text], return_labels=True)[0].split()\n", " assert(len(tokens) == len(punct_pred_1) == len(punct_pred_2))\n", " punct_tokens = []\n", " for t, p1, p2 in zip(tokens, punct_pred_1, punct_pred_2):\n", " token_formatted = t\n", " if p1[1] == \"U\":\n", " token_formatted = token_formatted.capitalize()\n", " punct_char = p1[0] if p1[0] != \"O\" else \"\"\n", " if punct_char == \"\" and p2[0] == \",\":\n", " punct_char = \",\"\n", " punct_tokens.append((token_formatted, punct_char))\n", " # use asr to fix punct and cap\n", " # TODO: improve this with alignment\n", " asr_out = whisper_predict(fp)\n", " asr_tokens = set(asr_out.split())\n", " asr_tokens_no_punct = set([t.strip(\".,?\") for t in asr_tokens])\n", " fixed_punct_tokens = []\n", " cap_next = False\n", " for t, p in punct_tokens:\n", " if cap_next:\n", " t = t.capitalize()\n", " cap_next = False\n", " if t + p not in asr_tokens and t + \"?\" in asr_tokens:\n", " p = \"?\"\n", " cap_next = True\n", " elif t + p not in asr_tokens and t + \".\" in asr_tokens:\n", " p = \".\"\n", " cap_next = True\n", " if t not in asr_tokens and t.upper() in asr_tokens_no_punct:\n", " t = t.upper()\n", " fixed_punct_tokens.append((t, p))\n", " text = \" \".join([\"\".join([t, p]) for t, p in fixed_punct_tokens])\n", " if False:\n", " \n", " # lowercase everything other than proper nouns\n", " # TODO: look at labels more properly and do alignment\n", " ner_preds = ner_predict(text)\n", " out_text = \" \".join([t if l != \"\" else t.lower() for t, l in ner_preds])\n", " # TODO: and/or exclude labels in ner?\n", " # final regex stuff\n", " out_text = re.sub(r\"\\bI\\b\", \"i\", out_text)\n", " out_text = re.sub(r\"(?<=[A-Za-z])\\*\\.\", \"*,\", out_text)\n", " out_text = re.sub(r\"\\<([A-Za-z]+?)\\>\", _lower, out_text)\n", " else:\n", " out_text = text\n", " out_tokens = out_text.split()\n", " adjusted_tokens = []\n", " for t1, t2 in zip(out_tokens[:-1], out_tokens[1:]):\n", " add_comma = False\n", " if t1.strip(\".,\") == t2.strip(\".,\"):\n", " add_comma = True\n", " if t1.strip(\".,\") in number_set and t2.strip(\".,\") in number_set:\n", " add_comma = True\n", " if add_comma:\n", " t = t1.strip(\".,\") + \",\"\n", " else:\n", " t = t1\n", " adjusted_tokens.append(t)\n", " adjusted_tokens.extend(out_tokens[-1:])\n", " adjusted_out_text = \" \".join(adjusted_tokens)\n", " out_texts.append(adjusted_out_text)\n", "out_str = \"\\n\".join([\"\\t\".join([fp, text]) for fp, text in zip(fp_list, out_texts)])" ] }, { "cell_type": "code", "execution_count": 35, "id": "c6b2e50f", "metadata": {}, "outputs": [], "source": [ "# with open(\"formatted_transcripts.tsv\", \"w\") as f:\n", "# f.write(out_str)\n", "with open(\"input_formatted_transcripts.tsv\", \"w\") as f:\n", " f.write(out_str)" ] }, { "cell_type": "code", "execution_count": 36, "id": "87ca3761", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[(0, 'Mustafa, do you have a second?'),\n", " (1, 'Yo, Andy, you got a second?'),\n", " (2, 'Seven, eight, nine.'),\n", " (3, 'Four, five, six.'),\n", " (4, 'Testing testing. Testing.'),\n", " (5,\n", " \"Yeah, sometimes I feel the same. Sometimes. Yeah, it's a little bit frustrating sometime.\"),\n", " (6,\n", " 'Yeah, well, sh*, should we try a design thing right now? like, test? After that we can talk about in the design world? '),\n", " (7, 'Was their fault to learn?'),\n", " (8, 'Andy, you wanna jump into a pairing session?'),\n", " (9,\n", " \"Yeah, and I feel like we all have could've all jumped into this thing and we all jumped off the thing, and we've been together each time.\"),\n", " (10, 'Is we shall we do?'),\n", " (11,\n", " \"I, I didn't notice that he was sick, but yeah, the I, I, I knew about the house. [not at my ti*]. It's funny because like sometimes like, uh, I don't know. Like the. The times are so different are so, so different of us working remotely that it's funny.\"),\n", " (12, \"Yep, what's up?\"),\n", " (13, 'Uh, not bad, not bad. How was your trip?'),\n", " (14, 'Lacking any neutral.'),\n", " (15,\n", " 'One, two, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen.'),\n", " (16, 'Testing speechly.'),\n", " (17, 'One two.'),\n", " (18, 'Testing one, two, three.'),\n", " (19, 'Testing.'),\n", " (20,\n", " \"Boom, it is really good. I don't feel like I've, I've you know, disconnected from a thing, and now I'm completely alone.\"),\n", " (21,\n", " 'Yo, I was just uh, debugging, uh, pop for pull flow, and we got it working. Uh, anyway, should we jump into the collaboration area?'),\n", " (22,\n", " 'Shall I ever question for you about how, how you want all the data from the feed back from *azur *am Because I think like the only option that I have is like sending back JSON.'),\n", " (23, 'Testing.'),\n", " (24,\n", " \"Yeah, exactly. A nice. I didn't hear your wave, but I saw your, your wave not so long ago. So that was successful, too.\"),\n", " (25, 'Problem, thank today.'),\n", " (26, 'Yeah, it feels more like work.'),\n", " (27, 'Sounds good. Sound good.'),\n", " (28, 'Yeah, I, just ah, this was fun. we*.'),\n", " (29, 'Yo, uh, you wanna jump on a one on one?'),\n", " (30, 'Yeah, gr*. thanks.'),\n", " (31, \"Test what's going on?\"),\n", " (32, 'Got it. Do you wanna pair on the Prs? Then we can get them out?'),\n", " (33, \"Hey, Jay, if you're around, Uh, I think I finished the profiles.\"),\n", " (34, 'Yep.'),\n", " (35,\n", " 'I know, and, and honestly, Speechly is, is really good, and how fast it is.'),\n", " (36,\n", " \"Uh, must to find out. we're just finishing our one on one, but we more often do. He was showing the Cam record. I saw you were here, so I figured we could do a quick short sync.\"),\n", " (37, 'You, Andy, quick question. Can you jump on a call?'),\n", " (38, \"Let's do it.\"),\n", " (39, 'Yeah.'),\n", " (40, 'Yo, Andy, you wanna try something new?'),\n", " (41, 'Yo, wanna see something cool?'),\n", " (42, 'Hello.'),\n", " (43, \"Oh, right, that's right. It's the funeral today.\"),\n", " (44, \"It says you're not in the walkie talkie area.\"),\n", " (45, 'Testing.'),\n", " (46,\n", " 'No, no, but I was just uh, watching a show on my computer with Pop on.'),\n", " (47,\n", " \"Yeah, I'm just uh, just finishing, uh, reviewing the changes right now.\"),\n", " (48, 'Great.'),\n", " (49,\n", " '*cup eh, do you wanna go to the collaboration area or in the one on ones?')]" ] }, "execution_count": 36, "metadata": {}, "output_type": "execute_result" } ], "source": [ "list(zip(range(len(out_texts)), out_texts))" ] }, { "cell_type": "code", "execution_count": 1, "id": "8668f3e3", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "/home/georg/notebooks/customers/speechly/punctuation\r\n" ] } ], "source": [ "!pwd" ] }, { "cell_type": "code", "execution_count": 3, "id": "02d0f72e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ " audios\t\t\t 'Punctuation guidelines.pdf' transcripts.tsv\r\n", " formatted_transcripts.tsv transcripts_base.tsv\t Untitled.ipynb\r\n" ] } ], "source": [ "!ls" ] }, { "cell_type": "code", "execution_count": 4, "id": "a0a424f3", "metadata": {}, "outputs": [], "source": [ "with open(\"/home/georg/notebooks/customers/speechly/punctuation/formatted_transcripts.tsv\") as f:\n", " out = f.read()" ] }, { "cell_type": "code", "execution_count": 9, "id": "0679913a", "metadata": {}, "outputs": [], "source": [ "with open(\"/home/georg/notebooks/customers/speechly/punctuation/transcripts_base.tsv\") as f:\n", " out = f.read()" ] }, { "cell_type": "code", "execution_count": 39, "id": "64510d17", "metadata": {}, "outputs": [], "source": [ "out = out.replace(\"audios/\", \"\").replace(\".wav\", \"\")" ] }, { "cell_type": "code", "execution_count": 41, "id": "e6e9caa2", "metadata": {}, "outputs": [], "source": [ "with open(\"/home/georg/notebooks/customers/speechly/punctuation/transcripts_base_2.tsv\", \"w\") as f:\n", " f.write(out)" ] }, { "cell_type": "code", "execution_count": 44, "id": "0a117246", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "511" ] }, "execution_count": 44, "metadata": {}, "output_type": "execute_result" } ], "source": [ "len(\" \".join([e.split(\"\\t\")[-1] for e in out.split(\"\\n\")]).split())" ] }, { "cell_type": "code", "execution_count": null, "id": "c6eed724", "metadata": {}, "outputs": [], "source": [ "50 segments, 511 words, 302 seconds" ] }, { "cell_type": "code", "execution_count": 61, "id": "3e4025b9", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "60913.90728476821" ] }, "execution_count": 61, "metadata": {}, "output_type": "execute_result" } ], "source": [ "10 * 60 * 60 / 302 * 511" ] }, { "cell_type": "code", "execution_count": null, "id": "468771ea", "metadata": {}, "outputs": [], "source": [ "6000 segments, 60k words" ] }, { "cell_type": "code", "execution_count": 50, "id": "fecdbd91", "metadata": {}, "outputs": [], "source": [ "import os\n", "d = \"/mnt/data-ssd-1/data/private/customer/speechly/pop.com/sample-50/audios\"\n", "fns = os.listdir(d)" ] }, { "cell_type": "code", "execution_count": 54, "id": "922dd46c", "metadata": {}, "outputs": [], "source": [ "from suno_utils.audio.conversion import get_duration_s" ] }, { "cell_type": "code", "execution_count": 55, "id": "2bcc1bb9", "metadata": {}, "outputs": [], "source": [ "l = []\n", "for fn in fns:\n", " fp = os.path.join(d, fn)\n", " l.append(get_duration_s(fp))" ] }, { "cell_type": "code", "execution_count": 58, "id": "c31ea96f", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "302.10600000000005" ] }, "execution_count": 58, "metadata": {}, "output_type": "execute_result" } ], "source": [ "np.sum(l)" ] }, { "cell_type": "code", "execution_count": 8, "id": "ff7621f8", "metadata": {}, "outputs": [], "source": [ "# with open(\"/home/georg/notebooks/customers/speechly/punctuation/transcripts_formatted.tsv\", \"w\") as f:\n", "# f.write(out)" ] }, { "cell_type": "code", "execution_count": null, "id": "eccf82bf", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "3cb157bd", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 112, "id": "7594cbff", "metadata": {}, "outputs": [], "source": [ "# fp = fp_list[4]\n", "# text = text_list[4]\n", "\n", "# tokens = normalize_whitespace(text).split()\n", "# # punctuate and capitalize\n", "# with suppress_logging():\n", "# with capture_output():\n", "# punct_pred_1 = punct_model.add_punctuation_capitalization([text], return_labels=True)[0].split()\n", "# punct_pred_2 = punct_2_model.add_punctuation_capitalization([text], return_labels=True)[0].split()\n", "# assert(len(tokens) == len(punct_pred_1) == len(punct_pred_2))\n", "# punct_tokens = []\n", "# for t, p1, p2 in zip(tokens, punct_pred_1, punct_pred_2):\n", "# token_formatted = t\n", "# if p1[1] == \"U\":\n", "# token_formatted = token_formatted.capitalize()\n", "# punct_char = p1[0] if p1[0] != \"O\" else \"\"\n", "# if punct_char == \"\" and p2[0] == \",\":\n", "# punct_char = \",\"\n", "# punct_tokens.append((token_formatted, punct_char))\n", "# # use asr to fix punct and cap\n", "# # TODO: improve this with alignment\n", "# asr_out = whisper_predict(fp)\n", "# asr_tokens = set(asr_out.split())\n", "# asr_tokens_no_punct = set([t.strip(\".,?\") for t in asr_tokens])\n", "# fixed_punct_tokens = []\n", "# cap_next = False\n", "# for t, p in punct_tokens:\n", "# if cap_next:\n", "# t = t.capitalize()\n", "# cap_next = False\n", "# if t + p not in asr_tokens and t + \"?\" in asr_tokens:\n", "# p = \"?\"\n", "# cap_next = True\n", "# elif t + p not in asr_tokens and t + \".\" in asr_tokens:\n", "# p = \".\"\n", "# cap_next = True\n", "# if t not in asr_tokens and t.upper() in asr_tokens_no_punct:\n", "# t = t.upper()\n", "# fixed_punct_tokens.append((t, p))\n", "# # lowercase everything other than proper nouns\n", "# # TODO: look at labels more properly and do alignment\n", "# ner_preds = ner_predict(\" \".join([\"\".join([t, p]) for t, p in fixed_punct_tokens]))\n", "# out_text = \" \".join([t if l != \"\" else t.lower() for t, l in ner_preds])\n", "# # TODO: and/or exclude labels in ner?\n", "# # final regex stuff\n", "# out_text = re.sub(r\"\\bI\\b\", \"i\", out_text)\n", "# out_tokens = out_text.split()\n", "# adjusted_tokens = []\n", "# for t1, t2 in zip(out_tokens[:-1], out_tokens[1:]):\n", "# add_comma = False\n", "# if t1.strip(\".,\") == t2.strip(\".,\"):\n", "# add_comma = True\n", "# if t1.strip(\".,\") in number_set and t2.strip(\".,\") in number_set:\n", "# add_comma = True\n", "# if add_comma:\n", "# t = t1.strip(\".,\") + \",\"\n", "# else:\n", "# t = t1\n", "# adjusted_tokens.append(t)\n", "# adjusted_tokens.extend(out_tokens[-1:])\n", "# adjusted_out_text = \" \".join(adjusted_tokens)\n", "# adjusted_out_text" ] }, { "cell_type": "code", "execution_count": null, "id": "e252e0a4", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "19f0c711", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "25f915fb", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "344a526e", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "0ae28f62", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "05811996", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "markdown", "id": "412d6bfd", "metadata": {}, "source": [ "## Playground" ] }, { "cell_type": "code", "execution_count": 2, "id": "acac965a", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2022-09-22 16:47:26 tokenize_and_classify:64] Creating ClassifyFst grammars.\n" ] } ], "source": [ "# create inverse text normalization instance\n", "from nemo_text_processing.inverse_text_normalization.inverse_normalize import InverseNormalizer\n", "inverse_normalizer = InverseNormalizer(lang='en')" ] }, { "cell_type": "code", "execution_count": 11, "id": "08398f23", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "we paid 15 for this desk\n" ] } ], "source": [ "spoken = \"we paid fifteen for this desk\"\n", "un_normalized = inverse_normalizer.inverse_normalize(spoken, verbose=False)\n", "print(un_normalized)" ] }, { "cell_type": "code", "execution_count": 8, "id": "e32f1009", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'we work for the f b i and others'" ] }, "execution_count": 8, "metadata": {}, "output_type": "execute_result" } ], "source": [ "spoken = \"we work for the f b i and others\"\n", "un_normalized = inverse_normalizer.inverse_normalize(spoken, verbose=False)\n", "un_normalized" ] }, { "cell_type": "code", "execution_count": 9, "id": "4298e937", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2022-09-22 11:29:03 punctuation_capitalization_model:1143] Using batch size 1 for inference\n", "[NeMo I 2022-09-22 11:29:03 punctuation_capitalization_infer_dataset:91] Max length: 11\n", "[NeMo I 2022-09-22 11:29:03 data_preprocessing:404] Some stats of the lengths of the sequences:\n", "[NeMo I 2022-09-22 11:29:03 data_preprocessing:406] Min: 9 | Max: 9 | Mean: 9.0 | Median: 9.0\n", "[NeMo I 2022-09-22 11:29:03 data_preprocessing:412] 75 percentile: 9.00\n", "[NeMo I 2022-09-22 11:29:03 data_preprocessing:413] 99 percentile: 9.00\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 55.93batch/s]\n" ] }, { "data": { "text/plain": [ "['We work for the F, B, I and others.']" ] }, "execution_count": 9, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model.add_punctuation_capitalization([\"we work for the f b i and others\"])" ] }, { "cell_type": "code", "execution_count": null, "id": "ae8b4dfd", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 19, "id": "471d20fa", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[NeMo I 2022-09-22 15:29:08 tokenize_and_classify:87] Creating ClassifyFst grammars.\n" ] } ], "source": [ "from nemo_text_processing.text_normalization.normalize import Normalizer\n", "normalizer = Normalizer(input_case='cased', lang='en')" ] }, { "cell_type": "code", "execution_count": 17, "id": "bd4649d0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Created cache_dir/en_tn_post_processing.far\n", "[NeMo I 2022-09-22 15:22:18 tokenize_and_classify_with_audio:101] Creating ClassifyFst grammars. This might take some time...\n", "Created cache_dir/_cased_en_tn_False_deterministic.far\n", "[NeMo I 2022-09-22 15:29:03 tokenize_and_classify_with_audio:229] ClassifyFst grammars are saved to cache_dir/_cased_en_tn_False_deterministic.far.\n", "Created cache_dir/en_tn_False_deterministic_verbalizer.far\n", "[NeMo I 2022-09-22 15:29:05 verbalize_final:76] VerbalizeFinalFst grammars are saved to cache_dir/en_tn_False_deterministic_verbalizer.far.\n" ] } ], "source": [ "from nemo_text_processing.text_normalization.normalize_with_audio import NormalizerWithAudio\n", "normalizer = NormalizerWithAudio(\n", " lang=\"en\",\n", " input_case=\"cased\",\n", " overwrite_cache=False,\n", " cache_dir=\"cache_dir\",\n", " )" ] }, { "cell_type": "code", "execution_count": 14, "id": "4d5edf70", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'we work for the FBI and others'" ] }, "execution_count": 14, "metadata": {}, "output_type": "execute_result" } ], "source": [ "written = \"we work for the FBI and others\"\n", "normalized = normalizer.normalize(written, verbose=False, punct_post_process=False)\n", "normalized" ] }, { "cell_type": "code", "execution_count": 20, "id": "460a93b4", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "we work for the FBI and others\n" ] }, { "data": { "text/plain": [ "{'we work for the F B I and others', 'we work for the FBI and others'}" ] }, "execution_count": 20, "metadata": {}, "output_type": "execute_result" } ], "source": [ "written = \"we work for the FBI and others\"\n", "normalized = normalizer_audio.normalize(written, n_tagged=10, punct_post_process=True)\n", "normalized" ] }, { "cell_type": "code", "execution_count": null, "id": "0ccfb347", "metadata": {}, "outputs": [], "source": [ "# import nemo\n", "# import nemo.collections.nlp as nemo_nlp\n", "# tagger = nemo_nlp.models.duplex_text_normalization.DuplexTaggerModel.from_pretrained(model_name=\"itn_en_t5\")\n", "# decoder = nemo_nlp.models.duplex_text_normalization.DuplexDecoderModel.from_pretrained(model_name=\"itn_en_t5\")\n", "# normalizer = nemo_nlp.models.duplex_text_normalization.DuplexTextNormalizationModel(tagger, decoder, lang=\"en\")" ] }, { "cell_type": "code", "execution_count": null, "id": "0d95035b", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "72ea4a6a", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "50325085", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "2e250ded", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "47727e3a", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "e16a24c6", "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 }