kazuk's picture
Duplicate from kazuk/youtube-whisper
c1cf3f5
raw
history blame
2.13 kB
import gradio as gr
import whisper
from pytube import YouTube
def get_audio(url):
yt = YouTube(url)
return yt.streams.filter(only_audio=True)[0].download(filename="tmp.mp4")
def get_transcript(url, model_size, lang, format):
model = whisper.load_model(model_size)
if lang == "None":
lang = None
result = model.transcribe(get_audio(url), fp16=False, language=lang)
if format == "None":
return result["text"]
elif format == ".srt":
return format_to_srt(result["segments"])
def format_to_srt(segments):
output = ""
for i, segment in enumerate(segments):
output += f"{i + 1}\n"
output += f"{format_timestamp(segment['start'])} --> {format_timestamp(segment['end'])}\n"
output += f"{segment['text']}\n\n"
return output
def format_timestamp(t):
hh = t//3600
mm = (t - hh*3600)//60
ss = t - hh*3600 - mm*60
mi = (t - int(t))*1000
return f"{int(hh):02d}:{int(mm):02d}:{int(ss):02d},{int(mi):03d}"
langs = ["None"] + sorted(list(whisper.tokenizer.LANGUAGES.values()))
model_size = list(whisper._MODELS.keys())
with gr.Blocks() as demo:
with gr.Row():
with gr.Column():
with gr.Row():
url = gr.Textbox(placeholder='Youtube video URL', label='URL')
with gr.Row():
model_size = gr.Dropdown(choices=model_size, value='tiny', label="Model")
lang = gr.Dropdown(choices=langs, value="None", label="Language (Optional)")
format = gr.Dropdown(choices=["None", ".srt"], value="None", label="Timestamps? (Optional)")
with gr.Row():
gr.Markdown("Larger models are more accurate, but slower. For 1min video, it'll take ~30s (tiny), ~1min (base), ~3min (small), ~5min (medium), etc.")
transcribe_btn = gr.Button('Transcribe')
with gr.Column():
outputs = gr.Textbox(placeholder='Transcription of the video', label='Transcription')
transcribe_btn.click(get_transcript, inputs=[url, model_size, lang, format], outputs=outputs)
demo.launch(debug=True)