Update: added script to handle data and transcribe.

2023-09-20_transcripts_Video-SRS.py

@@ -0,0 +1,14 @@
+from src.data_handling import download_audio, convert_to_wav
+from src.transcriber import transcribe_wav_file
+
+youtube_links = ['https://www.youtube.com/watch?v=BVVo1BaL6JA', 'https://www.youtube.com/watch?v=m0gr3snccVY&list=PL8p82vzFeTAj3cbHaXE9sC48pm4oZY2rH&index=2', 'https://www.youtube.com/watch?v=ThjgUfck8Sc&list=PL8p82vzFeTAj3cbHaXE9sC48pm4oZY2rH&index=3']
+#youtube_links = ['https://www.youtube.com/watch?v=6A5ZSiosgAk']
+
+for i, link in enumerate(youtube_links):
+    mp4_filename = f"audio_{i+1}.mp4"
+    wav_filename = f"audio_{i+1}.wav"
+
+    download_audio(link, mp4_filename)
+    convert_to_wav(mp4_filename, wav_filename)
+
+    transcribe_wav_file(wav_filename, i)

requirements.txt

@@ -0,0 +1,5 @@
+youtube-dl
+pfy
+pytube
+pydub
+git+https://github.com/m-bain/whisperx.git

src/__init__.py

@@ -0,0 +1 @@
+"""Functions to get video data and create transcripts as well as detailed information"""

src/data_handling.py

@@ -0,0 +1,13 @@
+def download_audio(youtube_link, filename):
+    yt = YouTube(youtube_link)
+
+    # get highest quality audio stream
+    audio_stream = yt.streams.get_audio_only()
+
+    # download
+    if not os.path.exists(filename):
+        audio_stream.download(filename=filename)
+
+def convert_to_wav(mp4_file, wav_file):
+    sound = AudioSegment.from_file(mp4_file, format="mp4")
+    sound.export(wav_file, format="wav")

src/transcriber.py

@@ -0,0 +1,40 @@
+google_colab_gpu = "cuda" if torch.cuda.is_available() else "cpu"
+
+def transcribe_wav_file(wav_filename, index):
+    batch_size = 16
+    compute_type = "float16"
+
+    model = whisperx.load_model("large-v2", device=google_colab_gpu, compute_type=compute_type)
+
+    audio = whisperx.load_audio(wav_filename)
+    result = model.transcribe(audio, batch_size=batch_size)
+
+    model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=google_colab_gpu)
+    result = whisperx.align(result["segments"], model_a, metadata, audio, google_colab_gpu, return_char_alignments=False)
+
+    # clear GPU memory
+    del model
+    del model_a
+    del metadata
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    diarize_model = whisperx.DiarizationPipeline(use_auth_token='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', device=google_colab_gpu)
+
+    diarize_segments = diarize_model(audio)
+    result = whisperx.assign_word_speakers(diarize_segments, result)
+
+    # clear GPU memory
+    del diarize_model
+    del diarize_segments
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    # Save full results to a json file
+    with open(f'result_{index+1}.json', 'w') as f:
+        json.dump(result, f)
+
+    # Save just the transcripts to a .txt file
+    transcripts = [seg['value'] for seg in result['segments']]
+    with open(f'transcripts_{index+1}.txt', 'w') as f:
+        f.write('\n'.join(transcripts))