Openais Whisperを使用してオーディオファイルを転写およびダイアリ化化する方法
Whisperは、Webから収集された680,000時間の多言語およびマルチタスク監視データでトレーニングされたOpenaiの最先端の音声認識システムです。この大きくて多様なデータセットは、アクセント、バックグラウンドノイズ、および専門言語に対する堅牢性の向上につながります。さらに、複数の言語での転写、およびそれらの言語から英語への翻訳が可能になります。 Openaiは、モデルとコードをリリースして、音声認識を活用する有用なアプリケーションを構築するための基盤として機能しました。
しかし、ささやきの大きな欠点は、会話で誰が話しているのかわからないということです。会話を分析するときは、それは問題です。これが日中化が始まる場所です。日記とは、会話で誰が話しているのかを特定するプロセスです。
このチュートリアルでは、スピーカーを識別する方法を学び、その後、ウィスパーの転写と一致させます。これを達成するためにpyannote-audioを使用します。始めましょう!
まず、オーディオファイルを準備する必要があります。 Lex Fridmansポッドキャストの最初の20分間のYannダウンロードを使用します。ビデオをダウンロードしてオーディオを抽出するには、 yt-dlpパッケージを使用します。
! pip install -U yt-dlpまた、FFMPEGをインストールする必要があります
! wget -O - -q https://github.com/yt-dlp/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-linux64-gpl.tar.xz | xz -qdc | tar -xこれで、コマンドラインを介して実際のダウンロードと音声抽出を行うことができます。
! yt-dlp -xv --ffmpeg-location ffmpeg-master-latest-linux64-gpl/bin --audio-format wav -o download.wav -- https://youtu.be/SGzMElJ11Ccこれで、ワーキングディレクトリにdownload.wavファイルがあります。オーディオの最初の20分をカットしましょう。このためにPydubパッケージを数行のコードで使用できます。
! pip install pydub from pydub import AudioSegment
t1 = 0 * 1000 # works in milliseconds
t2 = 20 * 60 * 1000
newAudio = AudioSegment . from_wav ( "download.wav" )
a = newAudio [ t1 : t2 ]
a . export ( "audio.wav" , format = "wav" ) audio.wavは、オーディオファイルの最初の20分です。
pyannote.audioは、スピーカーの珪化のためにPythonで書かれたオープンソースツールキットです。 Pytorch Machine Learningフレームワークに基づいて、スピーカーダイアリゼーションパイプラインを構築するために組み合わせて共同で最適化できるトレーニング可能なエンドツーエンドのニューラルビルディングブロックのセットを提供します。 pyannote.audioには、音声アクティビティ検出、スピーカーセグメンテーション、オーバーラップ音声検出、それらのほとんどの最先端のパフォーマンスに到達するスピーカーの埋め込みのための幅広いドメインをカバーする、優先されたモデルとパイプラインが付属しています。
Pyannoteをインストールし、ビデオオーディオで実行して、ダイアリゼーションを生成します。
! pip install pyannote.audio from pyannote . audio import Pipeline
pipeline = Pipeline . from_pretrained ( 'pyannote/speaker-diarization' ) DEMO_FILE = { 'uri' : 'blabal' , 'audio' : 'audio.wav' }
dz = pipeline ( DEMO_FILE )
with open ( "diarization.txt" , "w" ) as text_file :
text_file . write ( str ( dz ))これを印刷して、どのように見えるかを確認しましょう。
print(*list(dz.itertracks(yield_label = True))[:10], sep="n")
出力:
(<Segment(2.03344, 36.8128)>, 0, 'SPEAKER_00')
(<Segment(38.1122, 51.3759)>, 0, 'SPEAKER_00')
(<Segment(51.8653, 90.2053)>, 1, 'SPEAKER_01')
(<Segment(91.2853, 92.9391)>, 1, 'SPEAKER_01')
(<Segment(94.8628, 116.497)>, 0, 'SPEAKER_00')
(<Segment(116.497, 124.124)>, 1, 'SPEAKER_01')
(<Segment(124.192, 151.597)>, 1, 'SPEAKER_01')
(<Segment(152.018, 179.12)>, 1, 'SPEAKER_01')
(<Segment(180.318, 194.037)>, 1, 'SPEAKER_01')
(<Segment(195.016, 207.385)>, 0, 'SPEAKER_00')
これはすでにかなり良さそうですが、データを少しきれいにしましょう。
def millisec ( timeStr ):
spl = timeStr . split ( ":" )
s = ( int )(( int ( spl [ 0 ]) * 60 * 60 + int ( spl [ 1 ]) * 60 + float ( spl [ 2 ]) ) * 1000 )
return s
import re
dz = open ( 'diarization.txt' ). read (). splitlines ()
dzList = []
for l in dz :
start , end = tuple ( re . findall ( '[0-9]+:[0-9]+:[0-9]+.[0-9]+' , string = l ))
start = millisec ( start ) - spacermilli
end = millisec ( end ) - spacermilli
lex = not re . findall ( 'SPEAKER_01' , string = l )
dzList . append ([ start , end , lex ])
print ( * dzList [: 10 ], sep = ' n ' ) [33, 34812, True]
[36112, 49375, True]
[49865, 88205, False]
[89285, 90939, False]
[92862, 114496, True]
[114496, 122124, False]
[122191, 149596, False]
[150018, 177119, False]
[178317, 192037, False]
[193015, 205385, True]
これで、リストにダイアリゼーションデータがあります。最初の2つの数字は、ミリ秒単位でのスピーカーセグメントの開始時間と終了時間です。 3番目の数字は、スピーカーがlexであるかどうかを示すブール値です。
次に、デリミッターとしてスペーサーを使用して、ダイアリゼーションに従ってオーディオセグメントを添付します。
from pydub import AudioSegment
import re
sounds = spacer
segments = []
dz = open ( 'diarization.txt' ). read (). splitlines ()
for l in dz :
start , end = tuple ( re . findall ( '[0-9]+:[0-9]+:[0-9]+.[0-9]+' , string = l ))
start = int ( millisec ( start )) #milliseconds
end = int ( millisec ( end )) #milliseconds
segments . append ( len ( sounds ))
sounds = sounds . append ( audio [ start : end ], crossfade = 0 )
sounds = sounds . append ( spacer , crossfade = 0 )
sounds . export ( "dz.wav" , format = "wav" ) #Exports to a wav file in the current path. print ( segments [: 8 ])[2000, 38779, 54042, 94382, 98036, 121670, 131297, 160702]次に、ささやきを使用して、オーディオファイルのさまざまなセグメントを転写します。重要:Pyannote.audioとのバージョンの競合があり、エラーが発生します。私たちの回避策は、最初にpyannoteを実行してからささやきます。エラーを安全に無視できます。
Open AI Whisperのインストール。
! pip install git+https://github.com/openai/whisper.git 準備されたオーディオファイルでAIのささやきを開いています。転写をファイルに書き込みます。モデルサイズをニーズに合わせて調整できます。 GitHubのモデルカードにすべてのモデルを見つけることができます。
! whisper dz.wav --language en --model base [00:00.000 --> 00:04.720] The following is a conversation with Yann LeCun,
[00:04.720 --> 00:06.560] his second time on the podcast.
[00:06.560 --> 00:11.160] He is the chief AI scientist at Meta, formerly Facebook,
[00:11.160 --> 00:15.040] professor at NYU, touring award winner,
[00:15.040 --> 00:17.600] one of the seminal figures in the history
[00:17.600 --> 00:20.460] of machine learning and artificial intelligence,
...
.VTTファイルを使用するには、WebVTT-PYライブラリをインストールする必要があります。
! pip install -U webvtt-pyデータを見てみましょう。
import webvtt
captions = [[( int )( millisec ( caption . start )), ( int )( millisec ( caption . end )), caption . text ] for caption in webvtt . read ( 'dz.wav.vtt' )]
print ( * captions [: 8 ], sep = ' n ' ) [0, 4720, 'The following is a conversation with Yann LeCun,']
[4720, 6560, 'his second time on the podcast.']
[6560, 11160, 'He is the chief AI scientist at Meta, formerly Facebook,']
[11160, 15040, 'professor at NYU, touring award winner,']
[15040, 17600, 'one of the seminal figures in the history']
[17600, 20460, 'of machine learning and artificial intelligence,']
[20460, 23940, 'and someone who is brilliant and opinionated']
[23940, 25400, 'in the best kind of way,']
...
次に、各転写線をいくつかのダイアリゼーションに一致させ、HTMLファイルを生成してすべてを表示します。正しいタイミングを取得するには、ダイアリゼーションセグメントのない元のオーディオの部品の世話をする必要があります。オーディオの各セグメントに新しいDIVを追加します。
# we need this fore our HTML file (basicly just some styling)
preS = '<!DOCTYPE html>n<html lang="en">n <head>n <meta charset="UTF-8">n <meta name="viewport" content="width=device-width, initial-scale=1.0">n <meta http-equiv="X-UA-Compatible" content="ie=edge">n <title>Lexicap</title>n <style>n body {n font-family: sans-serif;n font-size: 18px;n color: #111;n padding: 0 0 1em 0;n }n .l {n color: #050;n }n .s {n display: inline-block;n }n .e {n display: inline-block;n }n .t {n display: inline-block;n }n #player {nttposition: sticky;ntttop: 20px;nttfloat: right;nt}n </style>n </head>n <body>n <h2>Yann LeCun: Dark Matter of Intelligence and Self-Supervised Learning | Lex Fridman Podcast #258</h2>n <div id="player"></div>n <script>n var tag = document.createElement('script');n tag.src = "https://www.youtube.com/iframe_api";n var firstScriptTag = document.getElementsByTagName('script')[0];n firstScriptTag.parentNode.insertBefore(tag, firstScriptTag);n var player;n function onYouTubeIframeAPIReady() {n player = new YT.Player('player', {n height: '210',n width: '340',n videoId: 'SGzMElJ11Cc',n });n }n function setCurrentTime(timepoint) {n player.seekTo(timepoint);n player.playVideo();n }n </script><br>n'
postS = 't</body>n</html>'
from datetime import timedelta
html = list(preS)
for i in range(len(segments)):
idx = 0
for idx in range(len(captions)):
if captions[idx][0] >= (segments[i] - spacermilli):
break;
while (idx < (len(captions))) and ((i == len(segments) - 1) or (captions[idx][1] < segments[i+1])):
c = captions[idx]
start = dzList[i][0] + (c[0] -segments[i])
if start < 0:
start = 0
idx += 1
start = start / 1000.0
startStr = '{0:02d}:{1:02d}:{2:02.2f}'.format((int)(start // 3600),
(int)(start % 3600 // 60),
start % 60)
html.append('ttt<div class="c">n')
html.append(f'tttt<a class="l" href="#{startStr}" id="{startStr}">link</a> |n')
html.append(f'tttt<div class="s"><a href="javascript:void(0);" onclick=setCurrentTime({int(start)})>{startStr}</a></div>n')
html.append(f'tttt<div class="t">{"[Lex]" if dzList[i][2] else "[Yann]"} {c[2]}</div>n')
html.append('ttt</div>nn')
html.append(postS)
s = "".join(html)
with open("lexicap.html", "w") as text_file:
text_file.write(s)
print(s)
Lablab Discordでは、このレポと人工知能に関連する他の多くのトピックについて説明します!今後の人工知能ハッカソンイベントをチェックアウトします