def __init__(self, model_id: str):
model_type = get_type(model_id)
if model_type is ModelType.ENCODERASR:
self.model = EncoderASR.from_hparams(source=model_id)
elif model_type is ModelType.ENCODERDECODERASR:
self.model = EncoderDecoderASR.from_hparams(source=model_id)
# Reduce latency
self.model.mods.decoder.beam_size = 1
else:
raise ValueError(
f"{model_type.value} is invalid for automatic-speech-recognition"
)
# Please define a `self.sampling_rate` for this pipeline
# to automatically read the input correctly
self.sampling_rate = self.model.hparams.sample_rate
def getTranscription(self):
asr_model = EncoderDecoderASR.from_hparams(
source="speechbrain/asr-crdnn-rnnlm-librispeech",
savedir="./pretrained_ASR")
transcription = asr_model.transcribe_file(self.path_to_wav)
return transcription.lower()
#with sr.Microphone() as source:
# stt = SpeechToText()
# r = sr.Recognizer()
# audio = r.listen(source, timeout=5)
# print("audio")
# print(audio)
# name = r.recognize_google(audio)
# print("google works, why can't you?")
# print(name)
# stt.saveAudio(audio)
# name = stt.getTranscription()
# print(name)
def call_huggingface(self, df):
assert self.model_url != '', "Error! A model URL is needed for HuggingFace scoring, but --asr_download_model is empty"
if self.tokenizer_url == '':
print(
f"Setting empty --tokenizer_url field identically to --asr_download_model: {self.model_url}"
)
self.tokenizer_url = self.model_url
if self.scoring_sorting == 'ascending':
df = df.sort_values(by=['n_frames']).reset_index(drop=True)
elif self.scoring_sorting == 'descending':
df = df.sort_values(by=['n_frames'],
ascending=False).reset_index(drop=True)
elif self.scoring_sorting == '':
pass
else:
raise NotImplementedError
print(f"Preparing dataloader for manifest {self.manifest}...")
dataset = AudioDataset(df)
dataloader = DataLoader(dataset,
batch_size=self.batch_size,
collate_fn=dataset.collater,
num_workers=self.num_workers,
pin_memory=True)
if self.hf_username == 'facebook':
print(f"Downloading tokenizer: {self.tokenizer_url}")
tokenizer = Wav2Vec2CTCTokenizer.from_pretrained(
self.tokenizer_url)
print(f"Downloading model: {self.model_url}")
model = Wav2Vec2ForCTC.from_pretrained(self.model_url)
elif self.hf_username == 'speechbrain':
if torch.cuda.is_available():
run_opts = {"device": "cuda"}
else:
run_opts = {"device": "cpu"}
print(f"Downloading model: {self.model_url}")
model = EncoderDecoderASR.from_hparams(source=self.model_url,
run_opts=run_opts,
savedir=os.path.join(
'pretrained_models',
self.hf_modelname))
else:
raise NotImplementedError
model.eval()
print("Scoring dataset...")
df['wer'] = np.nan
for batch in tqdm(dataloader):
indexes, waveforms, transcripts, wav_lens = batch
if self.hf_username == 'facebook':
output_logits = model(waveforms.squeeze()).logits
predicted_ids = torch.argmax(output_logits, dim=-1)
pred_transcripts = tokenizer.batch_decode(predicted_ids)
elif self.hf_username == 'speechbrain':
waveforms = waveforms.squeeze()
#waveforms = model.audio_normalizer(waveforms, self.sampling_rate)
pred_transcripts = model.transcribe_batch(waveforms,
wav_lens)[0]
for index, ref in enumerate(transcripts):
sample_id = indexes[index]
ref = transcripts[index]
pred = pred_transcripts[index]
measures = jiwer.compute_measures(ref, pred)
wer = measures['wer'] * 100.0
assert (
ref == df.loc[int(sample_id), 'tgt_text']
), "The reference text indicated by the sample ID in the transcripts file does not match with the one stored in the dataset!"
df.at[int(sample_id), 'wer'] = wer
return df
# !pip install librosa
import time
from time import perf_counter
import numpy as np
import matplotlib.pyplot as plt
import librosa
from pydub import AudioSegment
import os
from google.colab import files
import moviepy.editor
from transformers import pipeline
from speechbrain.pretrained import EncoderDecoderASR
asr_model2 = EncoderDecoderASR.from_hparams(source="speechbrain/asr-crdnn-rnnlm-librispeech", savedir="pretrained_models/asr-crdnn-rnnlm-librispeech")
def transcribe_audio(fileList = []):
listOfText = []
if fileList == [] :
uploaded = files.upload()
listOfAudios = list(uploaded.keys())
else:
listOfAudios = fileList
a = perf_counter()
for file in listOfAudios:
duration = librosa.get_duration(filename=file)
t1 = 0
t2 = duration * 1000 if duration < 30 else 30000
textTemp = ""
def load_asr_model():
asr_model = EncoderDecoderASR.from_hparams(
source="speechbrain/asr-transformer-transformerlm-librispeech",
savedir="pretrained_model/")
return asr_model
def asr_model():
"""Load model for the CTC segmentation test."""
asr_model = EncoderDecoderASR.from_hparams(
source="speechbrain/asr-transformer-transformerlm-librispeech")
return asr_model