Whisper

root=pretrained/whisper
mkdir -p ${root} && cd ${root}
git clone git@hf.co:openai/whisper-large-v3-turbo
cd ../..

pretrained/whisper/
└── whisper-large-v3-turbo

import einops
import torch
from transformers import AutomaticSpeechRecognitionPipeline, pipeline

from todd.utils import get_audio

assert torch.cuda.device_count() <= 4, (  # yapf: disable
    "Please use no more than 4 GPUs, in order to avoid RuntimeError."
)

url = (  # pylint: disable=invalid-name
    'https://github.com/SWivid/F5-TTS/raw/refs/heads/main/'
    'src/f5_tts/infer/examples/basic/basic_ref_zh.wav'
)
audio, _ = get_audio(url)
audio_array = audio.numpy()
audio_array = einops.rearrange(audio_array, '1 t -> t')

p: AutomaticSpeechRecognitionPipeline = pipeline(
    'automatic-speech-recognition',
    'pretrained/whisper/whisper-large-v3-turbo',
    torch_dtype='auto',
    device_map='auto',
)

result = p(audio_array)
print(result)

result = p(audio_array, generate_kwargs=dict(language='zh'))
print(result)

result = p(audio_array, generate_kwargs=dict(task='translate', language='en'))
print(result)

result = p(audio_array, return_timestamps=True)
print(result)

result = p(audio_array, return_timestamps='word')
print(result)