# if you want to use ASR model besides paraformer-bicif (like contextual paraformer)
|
# to get ASR results for long audio as well as timestamp prediction results,
|
# try this demo
|
from modelscope.pipelines import pipeline
|
from modelscope.utils.constant import Tasks
|
import os
|
import librosa
|
import soundfile as sf
|
|
param_dict = dict()
|
param_dict['hotword'] = "你的热词"
|
|
test_wav = 'YOUR_LONG_WAV.wav'
|
output_dir = './tmp'
|
os.system("mkdir -p {}".format(output_dir))
|
|
vad_pipeline = pipeline(
|
task=Tasks.voice_activity_detection,
|
model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
|
model_revision=None,
|
)
|
asr_pipeline = pipeline(
|
task=Tasks.auto_speech_recognition,
|
model="damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404",
|
output_dir=output_dir)
|
tp_pipeline = pipeline(
|
task=Tasks.speech_timestamp,
|
model='damo/speech_timestamp_prediction-v1-16k-offline',
|
output_dir=output_dir)
|
|
vad_res = vad_pipeline(audio_in=test_wav)
|
timestamps = vad_res['text']
|
|
samples = librosa.load(test_wav, sr=16000)[0]
|
wavseg_scp = "{}/wav.scp".format(output_dir)
|
|
with open(wavseg_scp, 'w') as fout:
|
for i, timestamp in enumerate(timestamps):
|
start = int(timestamp[0]/1000*16000)
|
end = int(timestamp[1]/1000*16000)
|
uttid = "wav_{}_{} ".format(start, end)
|
wavpath = '{}/wavseg_{}.wav'.format(output_dir, i)
|
_samples = samples[start:end]
|
sf.write(wavpath, _samples, 16000)
|
fout.write("{} {}\n".format(uttid, wavpath))
|
print("Wav segment done: {}".format(wavseg_scp))
|
|
asr_res = '{}/1best_recog/text'.format(output_dir)
|
tp_res = '{}/timestamp_prediction/tp_sync'.format(output_dir)
|
rec_result_asr = asr_pipeline(audio_in=wavseg_scp)
|
rec_result_tp = tp_pipeline(audio_in=wavseg_scp, text_in=asr_res)
|
print("Find your ASR results in {}, and timestamp prediction results in {}.".format(asr_res, tp_res))
|
