#!/usr/bin/env python3
|
# -*- encoding: utf-8 -*-
|
# Copyright FunASR (https://github.com/FunAudioLLM/SenseVoice). All Rights Reserved.
|
# MIT License (https://opensource.org/licenses/MIT)
|
|
import os
|
import torch
|
from pathlib import Path
|
from funasr import AutoModel
|
from funasr_onnx import SenseVoiceSmallONNX as SenseVoiceSmall
|
from funasr.utils.postprocess_utils import rich_transcription_postprocess
|
|
|
model_dir = "iic/SenseVoiceSmall"
|
model = AutoModel(
|
model=model_dir,
|
device="cuda:0",
|
)
|
|
res = model.export(type="onnx", quantize=False)
|
|
# export model init
|
model_path = "{}/.cache/modelscope/hub/{}".format(Path.home(), model_dir)
|
model_bin = SenseVoiceSmall(model_path)
|
|
# build tokenizer
|
try:
|
from funasr.tokenizer.sentencepiece_tokenizer import SentencepiecesTokenizer
|
tokenizer = SentencepiecesTokenizer(bpemodel=os.path.join(model_path, "chn_jpn_yue_eng_ko_spectok.bpe.model"))
|
except:
|
tokenizer = None
|
|
# inference
|
wav_or_scp = "/Users/shixian/Downloads/asr_example_hotword.wav"
|
language_list = [0]
|
textnorm_list = [15]
|
res = model_bin(wav_or_scp, language_list, textnorm_list, tokenizer=tokenizer)
|
print([rich_transcription_postprocess(i) for i in res])
|
