|
| 1 | +# %% |
| 2 | +# !pip install transformers==4.11.2 datasets soundfile sentencepiece torchaudio pyaudio |
| 3 | + |
| 4 | +# %% |
| 5 | +fromtransformersimport* |
| 6 | +importtorch |
| 7 | +importsoundfileassf |
| 8 | +# import librosa |
| 9 | +importos |
| 10 | +importtorchaudio |
| 11 | + |
| 12 | +# %% |
| 13 | +# model_name = "facebook/wav2vec2-base-960h" # 360MB |
| 14 | +model_name="facebook/wav2vec2-large-960h-lv60-self"# 1.18GB |
| 15 | + |
| 16 | +processor=Wav2Vec2Processor.from_pretrained(model_name) |
| 17 | +model=Wav2Vec2ForCTC.from_pretrained(model_name) |
| 18 | + |
| 19 | +# %% |
| 20 | +# audio_url = "http://www.fit.vutbr.cz/~motlicek/sympatex/f2bjrop1.0.wav" |
| 21 | +# audio_url = "http://www.fit.vutbr.cz/~motlicek/sympatex/f2bjrop1.1.wav" |
| 22 | +# audio_url = "http://www.fit.vutbr.cz/~motlicek/sympatex/f2btrop6.0.wav" |
| 23 | +# audio_url = "https://github.com/x4nth055/pythoncode-tutorials/raw/master/machine-learning/speech-recognition/16-122828-0002.wav" |
| 24 | +audio_url="https://github.com/x4nth055/pythoncode-tutorials/raw/master/machine-learning/speech-recognition/30-4447-0004.wav" |
| 25 | +# audio_url = "https://github.com/x4nth055/pythoncode-tutorials/raw/master/machine-learning/speech-recognition/7601-291468-0006.wav" |
| 26 | +# audio_url = "https://file-examples-com.github.io/uploads/2017/11/file_example_WAV_1MG.wav" |
| 27 | +# audio_url = "http://www0.cs.ucl.ac.uk/teaching/GZ05/samples/lathe.wav" |
| 28 | + |
| 29 | +# %% |
| 30 | +# load our wav file |
| 31 | +speech,sr=torchaudio.load(audio_url) |
| 32 | +speech=speech.squeeze() |
| 33 | +# or using librosa |
| 34 | +# speech, sr = librosa.load(audio_file, sr=16000) |
| 35 | +sr,speech.shape |
| 36 | + |
| 37 | +# %% |
| 38 | +# resample from whatever the audio sampling rate to 16000 |
| 39 | +resampler=torchaudio.transforms.Resample(sr,16000) |
| 40 | +speech=resampler(speech) |
| 41 | +speech.shape |
| 42 | + |
| 43 | +# %% |
| 44 | +# tokenize our wav |
| 45 | +input_values=processor(speech,return_tensors="pt",sampling_rate=16000)["input_values"] |
| 46 | +input_values.shape |
| 47 | + |
| 48 | +# %% |
| 49 | +# perform inference |
| 50 | +logits=model(input_values)["logits"] |
| 51 | +logits.shape |
| 52 | + |
| 53 | +# %% |
| 54 | +# use argmax to get the predicted IDs |
| 55 | +predicted_ids=torch.argmax(logits,dim=-1) |
| 56 | +predicted_ids.shape |
| 57 | + |
| 58 | +# %% |
| 59 | +# decode the IDs to text |
| 60 | +transcription=processor.decode(predicted_ids[0]) |
| 61 | +transcription.lower() |
| 62 | + |
| 63 | +# %% |
| 64 | +defget_transcription(audio_path): |
| 65 | +# load our wav file |
| 66 | +speech,sr=torchaudio.load(audio_path) |
| 67 | +speech=speech.squeeze() |
| 68 | +# or using librosa |
| 69 | +# speech, sr = librosa.load(audio_file, sr=16000) |
| 70 | +# resample from whatever the audio sampling rate to 16000 |
| 71 | +resampler=torchaudio.transforms.Resample(sr,16000) |
| 72 | +speech=resampler(speech) |
| 73 | +# tokenize our wav |
| 74 | +input_values=processor(speech,return_tensors="pt",sampling_rate=16000)["input_values"] |
| 75 | +# perform inference |
| 76 | +logits=model(input_values)["logits"] |
| 77 | +# use argmax to get the predicted IDs |
| 78 | +predicted_ids=torch.argmax(logits,dim=-1) |
| 79 | +# decode the IDs to text |
| 80 | +transcription=processor.decode(predicted_ids[0]) |
| 81 | +returntranscription.lower() |
| 82 | + |
| 83 | +# %% |
| 84 | +get_transcription(audio_url) |
| 85 | + |
| 86 | +# %% |
| 87 | +importpyaudio |
| 88 | +importwave |
| 89 | + |
| 90 | +# the file name output you want to record into |
| 91 | +filename="recorded.wav" |
| 92 | +# set the chunk size of 1024 samples |
| 93 | +chunk=1024 |
| 94 | +# sample format |
| 95 | +FORMAT=pyaudio.paInt16 |
| 96 | +# mono, change to 2 if you want stereo |
| 97 | +channels=1 |
| 98 | +# 44100 samples per second |
| 99 | +sample_rate=16000 |
| 100 | +record_seconds=10 |
| 101 | +# initialize PyAudio object |
| 102 | +p=pyaudio.PyAudio() |
| 103 | +# open stream object as input & output |
| 104 | +stream=p.open(format=FORMAT, |
| 105 | +channels=channels, |
| 106 | +rate=sample_rate, |
| 107 | +input=True, |
| 108 | +output=True, |
| 109 | +frames_per_buffer=chunk) |
| 110 | +frames= [] |
| 111 | +print("Recording...") |
| 112 | +foriinrange(int(sample_rate/chunk*record_seconds)): |
| 113 | +data=stream.read(chunk) |
| 114 | +# if you want to hear your voice while recording |
| 115 | +# stream.write(data) |
| 116 | +frames.append(data) |
| 117 | +print("Finished recording.") |
| 118 | +# stop and close stream |
| 119 | +stream.stop_stream() |
| 120 | +stream.close() |
| 121 | +# terminate pyaudio object |
| 122 | +p.terminate() |
| 123 | +# save audio file |
| 124 | +# open the file in 'write bytes' mode |
| 125 | +wf=wave.open(filename,"wb") |
| 126 | +# set the channels |
| 127 | +wf.setnchannels(channels) |
| 128 | +# set the sample format |
| 129 | +wf.setsampwidth(p.get_sample_size(FORMAT)) |
| 130 | +# set the sample rate |
| 131 | +wf.setframerate(sample_rate) |
| 132 | +# write the frames as bytes |
| 133 | +wf.writeframes(b"".join(frames)) |
| 134 | +# close the file |
| 135 | +wf.close() |
| 136 | + |
| 137 | +# %% |
| 138 | +get_transcription("recorded.wav") |
| 139 | + |
| 140 | +# %% |
| 141 | + |
| 142 | + |
| 143 | + |