pred_ppg_c

[EmreOzkose]

Hi, thank you for sharing code.

In whisperconvert_longaudio , there is from preprocess_ppg import pred_ppg_c, but actually there is not any pred_ppg_c function. Can you share that ?

[moonnee]

hi, you can add codes below to preprocess_ppg.py. I find this works.

def pred_ppg_c(whisper: Whisper, wavPath):
    audio, sr = librosa.load(wavPath,sr=None)
    if len(audio) >= sr * 29:
        print(wavPath,"cut to 29s")
        audio = audio[:sr * 29]
        #librosa.output.write_wav("your_audio_file.wav", audio, sr)
        sf.write(wavPath, audio, sr)
    audio = load_audio(wavPath)
    audln = audio.shape[0]
    ppgln = audln // 320
    # audio = pad_or_trim(audio)
    mel = log_mel_spectrogram(audio).to(whisper.device)
    with torch.no_grad():
        ppg = whisper.encoder(mel.unsqueeze(0)).squeeze().data.cpu().float().numpy()
        
        if ppgln>ppg.shape[0]:
            print("ppgln>ppg.shape[0]")
        ppg = ppg[:ppgln,] # [length, dim=1024]
        #if audln // 320<ppg.shape[0]:
        #    print("audln // 320<ppg.shape[0]")
    return ppg

[EmreOzkose]

thank you