whisper-demo-german / run_demo_hf_multiple_models.py
bofenghuang's picture
up
8ecd0fd
import logging
import warnings
import gradio as gr
import pytube as pt
import torch
from huggingface_hub import model_info
from transformers import pipeline
from transformers.utils.logging import disable_progress_bar
warnings.filterwarnings("ignore")
disable_progress_bar()
DEFAULT_MODEL_NAME = "bofenghuang/whisper-large-v2-cv11-german"
# make sure no OOM
MODEL_NAMES = [
"bofenghuang/whisper-medium-cv11-german",
"bofenghuang/whisper-large-v2-cv11-german",
]
LANG = "de"
CHUNK_LENGTH_S = 30
MAX_NEW_TOKENS = 225
logging.basicConfig(
format="%(asctime)s [%(levelname)s] [%(name)s] %(message)s",
datefmt="%Y-%m-%dT%H:%M:%SZ",
)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)
device = 0 if torch.cuda.is_available() else "cpu"
logger.info(f"Model will be loaded on device `{device}`")
cached_models = {}
def print_cuda_memory_info():
used_mem, tot_mem = torch.cuda.mem_get_info()
logger.info(f"CUDA memory info - Free: {used_mem / 1024 ** 3:.2f} Gb, used: {(tot_mem - used_mem) / 1024 ** 3:.2f} Gb, total: {tot_mem / 1024 ** 3:.2f} Gb")
def print_memory_info():
# todo
if device == "cpu":
pass
else:
print_cuda_memory_info()
def maybe_load_cached_pipeline(model_name):
pipe = cached_models.get(model_name)
if pipe is None:
# load pipeline
# todo: set decoding option for pipeline
pipe = pipeline(
task="automatic-speech-recognition",
model=model_name,
chunk_length_s=CHUNK_LENGTH_S,
device=device,
)
# set forced_decoder_ids
pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language=LANG, task="transcribe")
# limit genneration max length
pipe.model.config.max_length = MAX_NEW_TOKENS + 1
logger.info(f"`{model_name}` pipeline has been initialized")
print_memory_info()
cached_models[model_name] = pipe
return pipe
def transcribe(microphone, file_upload, model_name):
warn_output = ""
if (microphone is not None) and (file_upload is not None):
warn_output = (
"WARNING: You've uploaded an audio file and used the microphone. "
"The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
)
elif (microphone is None) and (file_upload is None):
return "ERROR: You have to either use the microphone or upload an audio file"
file = microphone if microphone is not None else file_upload
pipe = maybe_load_cached_pipeline(model_name)
text = pipe(file)["text"]
logger.info(f"Transcription by `{model_name}`: {text}")
return warn_output + text
def _return_yt_html_embed(yt_url):
video_id = yt_url.split("?v=")[-1]
HTML_str = (
f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
" </center>"
)
return HTML_str
def yt_transcribe(yt_url, model_name):
yt = pt.YouTube(yt_url)
html_embed_str = _return_yt_html_embed(yt_url)
stream = yt.streams.filter(only_audio=True)[0]
stream.download(filename="audio.mp3")
pipe = maybe_load_cached_pipeline(model_name)
text = pipe("audio.mp3")["text"]
logger.info(f"Transcription: {text}")
return html_embed_str, text
# load default model
maybe_load_cached_pipeline(DEFAULT_MODEL_NAME)
demo = gr.Blocks()
mf_transcribe = gr.Interface(
fn=transcribe,
inputs=[
gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Record"),
gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Upload File"),
gr.inputs.Dropdown(choices=MODEL_NAMES, default=DEFAULT_MODEL_NAME, label="Whisper Model"),
],
# outputs="text",
outputs=gr.outputs.Textbox(label="Transcription"),
layout="horizontal",
theme="huggingface",
title="Whisper German Demo 🇩🇪 : Transcribe Audio",
description="Transcribe long-form microphone or audio inputs with the click of a button!",
allow_flagging="never",
)
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.Dropdown(choices=MODEL_NAMES, default=DEFAULT_MODEL_NAME, label="Whisper Model"),
],
# outputs=["html", "text"],
outputs=[
gr.outputs.HTML(label="YouTube Page"),
gr.outputs.Textbox(label="Transcription"),
],
layout="horizontal",
theme="huggingface",
title="Whisper German Demo 🇩🇪 : Transcribe YouTube",
description="Transcribe long-form YouTube videos with the click of a button!",
allow_flagging="never",
)
with demo:
gr.TabbedInterface([mf_transcribe, yt_transcribe], ["Transcribe Audio", "Transcribe YouTube"])
# demo.launch(server_name="0.0.0.0", debug=True, share=True)
demo.launch(enable_queue=True)