asahi417 commited on
Commit
d78252d
β€’
1 Parent(s): fba42a4

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +11 -51
app.py CHANGED
@@ -14,7 +14,6 @@ FILE_LIMIT_MB = 1000
14
  YT_LENGTH_LIMIT_S = 3600 # limit to 1 hour YouTube files
15
 
16
  device = 0 if torch.cuda.is_available() else "cpu"
17
-
18
  pipe = pipeline(
19
  task="automatic-speech-recognition",
20
  model=MODEL_NAME,
@@ -23,47 +22,35 @@ pipe = pipeline(
23
  )
24
 
25
 
26
- def transcribe(inputs, task):
27
  if inputs is None:
28
  raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
29
-
30
- text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
31
- return text
32
 
33
 
34
  def _return_yt_html_embed(yt_url):
35
  video_id = yt_url.split("?v=")[-1]
36
- HTML_str = (
37
- f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
38
- " </center>"
39
- )
40
- return HTML_str
41
 
42
  def download_yt_audio(yt_url, filename):
43
  info_loader = youtube_dl.YoutubeDL()
44
-
45
  try:
46
  info = info_loader.extract_info(yt_url, download=False)
47
  except youtube_dl.utils.DownloadError as err:
48
  raise gr.Error(str(err))
49
-
50
  file_length = info["duration_string"]
51
  file_h_m_s = file_length.split(":")
52
  file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]
53
-
54
  if len(file_h_m_s) == 1:
55
  file_h_m_s.insert(0, 0)
56
  if len(file_h_m_s) == 2:
57
  file_h_m_s.insert(0, 0)
58
  file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]
59
-
60
  if file_length_s > YT_LENGTH_LIMIT_S:
61
  yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
62
  file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
63
  raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")
64
-
65
  ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
66
-
67
  with youtube_dl.YoutubeDL(ydl_opts) as ydl:
68
  try:
69
  ydl.download([yt_url])
@@ -71,76 +58,49 @@ def download_yt_audio(yt_url, filename):
71
  raise gr.Error(str(err))
72
 
73
 
74
- def yt_transcribe(yt_url, task, max_filesize=75.0):
75
  html_embed_str = _return_yt_html_embed(yt_url)
76
-
77
  with tempfile.TemporaryDirectory() as tmpdirname:
78
  filepath = os.path.join(tmpdirname, "video.mp4")
79
  download_yt_audio(yt_url, filepath)
80
  with open(filepath, "rb") as f:
81
  inputs = f.read()
82
-
83
  inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
84
  inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
85
-
86
- text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
87
-
88
  return html_embed_str, text
89
 
90
 
91
  demo = gr.Blocks()
92
-
93
  mf_transcribe = gr.Interface(
94
  fn=transcribe,
95
- inputs=[
96
- gr.components.Audio(sources=["microphone"], type="filepath"),
97
- gr.components.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
98
- ],
99
  outputs="text",
100
  layout="horizontal",
101
  theme="huggingface",
102
  title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
103
- description=(
104
- "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the Japanese Whisper"
105
- f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files"
106
- " of arbitrary length."
107
- ),
108
  allow_flagging="never",
109
  )
110
 
111
  file_transcribe = gr.Interface(
112
  fn=transcribe,
113
- inputs=[
114
- gr.components.Audio(sources=["upload"], type="filepath", label="Audio file"),
115
- gr.components.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
116
- ],
117
  outputs="text",
118
  layout="horizontal",
119
  theme="huggingface",
120
  title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
121
- description=(
122
- "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the OpenAI Whisper"
123
- f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files"
124
- " of arbitrary length."
125
- ),
126
  allow_flagging="never",
127
  )
128
-
129
  yt_transcribe = gr.Interface(
130
  fn=yt_transcribe,
131
- inputs=[
132
- gr.components.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
133
- gr.components.Radio(["transcribe", "translate"], label="Task", default="transcribe")
134
- ],
135
  outputs=["html", "text"],
136
  layout="horizontal",
137
  theme="huggingface",
138
  title="Whisper Large V3: Transcribe YouTube",
139
- description=(
140
- "Transcribe long-form YouTube videos with the click of a button! Demo uses the OpenAI Whisper checkpoint"
141
- f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe video files of"
142
- " arbitrary length."
143
- ),
144
  allow_flagging="never",
145
  )
146
 
 
14
  YT_LENGTH_LIMIT_S = 3600 # limit to 1 hour YouTube files
15
 
16
  device = 0 if torch.cuda.is_available() else "cpu"
 
17
  pipe = pipeline(
18
  task="automatic-speech-recognition",
19
  model=MODEL_NAME,
 
22
  )
23
 
24
 
25
+ def transcribe(inputs):
26
  if inputs is None:
27
  raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
28
+ return pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
 
 
29
 
30
 
31
  def _return_yt_html_embed(yt_url):
32
  video_id = yt_url.split("?v=")[-1]
33
+ return f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe> </center>'
 
 
 
 
34
 
35
  def download_yt_audio(yt_url, filename):
36
  info_loader = youtube_dl.YoutubeDL()
 
37
  try:
38
  info = info_loader.extract_info(yt_url, download=False)
39
  except youtube_dl.utils.DownloadError as err:
40
  raise gr.Error(str(err))
 
41
  file_length = info["duration_string"]
42
  file_h_m_s = file_length.split(":")
43
  file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]
 
44
  if len(file_h_m_s) == 1:
45
  file_h_m_s.insert(0, 0)
46
  if len(file_h_m_s) == 2:
47
  file_h_m_s.insert(0, 0)
48
  file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]
 
49
  if file_length_s > YT_LENGTH_LIMIT_S:
50
  yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
51
  file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
52
  raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")
 
53
  ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
 
54
  with youtube_dl.YoutubeDL(ydl_opts) as ydl:
55
  try:
56
  ydl.download([yt_url])
 
58
  raise gr.Error(str(err))
59
 
60
 
61
+ def yt_transcribe(yt_url, max_filesize=75.0):
62
  html_embed_str = _return_yt_html_embed(yt_url)
 
63
  with tempfile.TemporaryDirectory() as tmpdirname:
64
  filepath = os.path.join(tmpdirname, "video.mp4")
65
  download_yt_audio(yt_url, filepath)
66
  with open(filepath, "rb") as f:
67
  inputs = f.read()
 
68
  inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
69
  inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
70
+ text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
 
 
71
  return html_embed_str, text
72
 
73
 
74
  demo = gr.Blocks()
 
75
  mf_transcribe = gr.Interface(
76
  fn=transcribe,
77
+ inputs=[gr.components.Audio(sources=["microphone"], type="filepath")],
 
 
 
78
  outputs="text",
79
  layout="horizontal",
80
  theme="huggingface",
81
  title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
82
+ description=f"Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the Japanese Whisper checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files of arbitrary length.",
 
 
 
 
83
  allow_flagging="never",
84
  )
85
 
86
  file_transcribe = gr.Interface(
87
  fn=transcribe,
88
+ inputs=[gr.components.Audio(sources=["upload"], type="filepath", label="Audio file")],
 
 
 
89
  outputs="text",
90
  layout="horizontal",
91
  theme="huggingface",
92
  title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
93
+ description=f"Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the OpenAI Whisper checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe audio files of arbitrary length.",
 
 
 
 
94
  allow_flagging="never",
95
  )
 
96
  yt_transcribe = gr.Interface(
97
  fn=yt_transcribe,
98
+ inputs=[gr.components.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")],
 
 
 
99
  outputs=["html", "text"],
100
  layout="horizontal",
101
  theme="huggingface",
102
  title="Whisper Large V3: Transcribe YouTube",
103
+ description=f"Transcribe long-form YouTube videos with the click of a button! Demo uses the OpenAI Whisper checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and πŸ€— Transformers to transcribe video files of arbitrary length.",
 
 
 
 
104
  allow_flagging="never",
105
  )
106