fix_sampling_rate

app.py

@@ -1,85 +1,37 @@
-import os
-import time
-import tempfile
-from math import floor
-from typing import Optional, List, Dict, Any
-
 import torch
+
 import gradio as gr
 import yt_dlp as youtube_dl
 from transformers import pipeline
 from transformers.pipelines.audio_utils import ffmpeg_read
 
+import tempfile
+import os
 
-# configuration
-MODEL_NAME = "kotoba-tech/kotoba-whisper-v1.1"
-BATCH_SIZE = 16
-CHUNK_LENGTH_S = 15
+MODEL_NAME = "kotoba-tech/kotoba-whisper-v1.0"
+BATCH_SIZE = 8
 FILE_LIMIT_MB = 1000
 YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files
-# device setting
-if torch.cuda.is_available():
-    torch_dtype = torch.bfloat16
-    device = "cuda:0"
-    model_kwargs = {'attn_implementation': 'sdpa'}
-else:
-    torch_dtype = torch.float32
-    device = "cpu"
-    model_kwargs = {}
-# define the pipeline
+
+device = 0 if torch.cuda.is_available() else "cpu"
 pipe = pipeline(
+    task="automatic-speech-recognition",
     model=MODEL_NAME,
-    chunk_length_s=CHUNK_LENGTH_S,
-    batch_size=BATCH_SIZE,
-    torch_dtype=torch_dtype,
+    chunk_length_s=30,
     device=device,
-    model_kwargs=model_kwargs,
-    trust_remote_code=True
 )
 
 
-def format_time(start: Optional[float], end: Optional[float]):
-
-    def _format_time(seconds: Optional[float]):
-        if seconds is None:
-            return "complete    "
-        minutes = floor(seconds / 60)
-        hours = floor(seconds / 3600)
-        seconds = seconds - hours * 3600 - minutes * 60
-        m_seconds = floor(round(seconds - floor(seconds), 3) * 10 ** 3)
-        seconds = floor(seconds)
-        return f'{hours:02}:{minutes:02}:{seconds:02}.{m_seconds:03}'
-
-    return f"[{_format_time(start)}-> {_format_time(end)}]:"
-
-
-def get_prediction(inputs, prompt: Optional[str]):
-    generate_kwargs = {"language": "japanese", "task": "transcribe"}
-    if prompt:
-        generate_kwargs['prompt_ids'] = pipe.tokenizer.get_prompt_ids(prompt, return_tensors='pt').to(device)
-    prediction = pipe(inputs, return_timestamps=True, generate_kwargs=generate_kwargs)
-    text = "".join([c['text'] for c in prediction['chunks']])
-    text_timestamped = "\n".join([
-        f"{format_time(*c['timestamp'])} {c['text']}" for c in prediction['chunks']
-    ])
-    return text, text_timestamped
-
-
-def transcribe(inputs: str, prompt):
+def transcribe(inputs):
     if inputs is None:
         raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
-    with open(inputs, "rb") as f:
-        inputs = f.read()
-    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
-    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
-    return get_prediction(inputs, prompt)
+    return pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
 
 
 def _return_yt_html_embed(yt_url):
     video_id = yt_url.split("?v=")[-1]
     return f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe> </center>'
 
-
 def download_yt_audio(yt_url, filename):
     info_loader = youtube_dl.YoutubeDL()
     try:
@@ -106,7 +58,7 @@ def download_yt_audio(yt_url, filename):
             raise gr.Error(str(err))
 
 
-def yt_transcribe(yt_url, prompt):
+def yt_transcribe(yt_url, max_filesize=75.0):
     html_embed_str = _return_yt_html_embed(yt_url)
     with tempfile.TemporaryDirectory() as tmpdirname:
         filepath = os.path.join(tmpdirname, "video.mp4")
@@ -115,18 +67,15 @@ def yt_transcribe(yt_url, prompt):
             inputs = f.read()
     inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
     inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}
-    text, text_timestamped = get_prediction(inputs, prompt)
-    return html_embed_str, text, text_timestamped
+    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": "transcribe"}, return_timestamps=True)["text"]
+    return html_embed_str, text
 
 
 demo = gr.Blocks()
 mf_transcribe = gr.Interface(
     fn=transcribe,
-    inputs=[
-        gr.inputs.Audio(source="microphone", type="filepath", optional=True),
-        gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
-    ],
-    outputs=["text", "text"],
+    inputs=[gr.inputs.Audio(source="microphone", type="filepath", optional=True)],
+    outputs="text",
     layout="horizontal",
     theme="huggingface",
     title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
@@ -136,11 +85,8 @@ mf_transcribe = gr.Interface(
 
 file_transcribe = gr.Interface(
     fn=transcribe,
-    inputs=[
-        gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Audio file"),
-        gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
-    ],
-    outputs=["text", "text"],
+    inputs=[gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Audio file")],
+    outputs="text",
     layout="horizontal",
     theme="huggingface",
     title=f"Transcribe Audio with {os.path.basename(MODEL_NAME)}",
@@ -149,11 +95,8 @@ file_transcribe = gr.Interface(
 )
 yt_transcribe = gr.Interface(
     fn=yt_transcribe,
-    inputs=[
-        gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
-        gr.inputs.Textbox(lines=1, placeholder="Prompt", optional=True),
-    ],
-    outputs=["html", "text", "text"],
+    inputs=[gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")],
+    outputs=["html", "text"],
     layout="horizontal",
     theme="huggingface",
     title=f"Transcribe YouTube with {os.path.basename(MODEL_NAME)}",

requirements.txt

@@ -1,5 +1,3 @@
 git+https://github.com/huggingface/transformers
 torch
-yt-dlp
-punctuators==0.0.5
-stable-ts==2.16.0
+yt-dlp