whisper.py

from pyannote.audio import Pipeline
from pydub import AudioSegment
import os
from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torchaudio
import torch

device = 0 if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float32


MODEL_NAME = "openai/whisper-large-v3"
model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype).to(device)
processor = WhisperProcessor.from_pretrained(MODEL_NAME)
pipeline_vad = Pipeline.from_pretrained("pyannote/voice-activity-detection", use_auth_token=os.environ.get("HF_TOKEN"))
threshold = 15000  # adjust max duration threshold
segments_dir = "."

def clean_text(input_text):
    
    remove_chars = ['.', ',', ';', ':', '¿', '?', '«', '»', '-', '¡', '!', '@',
                     '*', '{', '}', '[', ']', '=', '/', '\\', '&', '#', '…']

    output_text = ''.join(char if char not in remove_chars else ' ' for char in input_text) #removing special chars
    return (' '.join(output_text.split()).lower()) #remove extra spaces and return cleaned text

def convert_forced_to_tokens(forced_decoder_ids):
    forced_decoder_tokens = []
    for i, (idx, token) in enumerate(forced_decoder_ids):
        if token is not None:
            forced_decoder_tokens.append([idx, processor.tokenizer.decode(token)])
        else:
            forced_decoder_tokens.append([idx, token])
    return forced_decoder_tokens

def generate_1st_chunk(audio):

    input_audio, sample_rate = torchaudio.load(audio)
    input_audio = torchaudio.transforms.Resample(sample_rate, 16000)(input_audio)
    
    input_speech = input_audio[0]

    input_features = processor(input_speech, 
                                    sampling_rate=16_000, 
                                    return_tensors="pt", torch_dtype=torch_dtype).input_features.to(device)

    forced_decoder_ids = []
    forced_decoder_ids.append([1,50270]) #[1, '<|ca|>']
    forced_decoder_ids.append([2,50262]) #[2, '<|es|>']
    forced_decoder_ids.append([3,50360]) #[3, '<|transcribe|>']

    forced_decoder_ids_modified = forced_decoder_ids

    # we need to force these tokens
    forced_decoder_ids = []

    # now we need to append the prefix tokens (lang, task, timestamps)
    offset = len(forced_decoder_ids)
    for idx, token in forced_decoder_ids_modified:
        forced_decoder_ids.append([idx + offset , token])
    
    model.generation_config.forced_decoder_ids = forced_decoder_ids

    pred_ids = model.generate(input_features, 
                                    return_timestamps=True,
                                    max_new_tokens=128)
    #exclude prompt from output
    forced_decoder_tokens = convert_forced_to_tokens(forced_decoder_ids)
    output = processor.decode(pred_ids[0][len(forced_decoder_tokens) + 1:], skip_special_tokens=True)
    output_tokens = processor.batch_decode(pred_ids, skip_special_tokens=False)

    return output[1:]

def generate_from_2nd_chunk(audio, prev_prompt):

    input_audio, sample_rate = torchaudio.load(audio)
    input_audio = torchaudio.transforms.Resample(sample_rate, 16000)(input_audio)
    
    input_speech = input_audio[0]

    input_features = processor(input_speech, 
                                    sampling_rate=16_000, 
                                    return_tensors="pt", torch_dtype=torch_dtype).input_features.to(device)
    forced_decoder_ids = []

    forced_decoder_ids.append([1,50270]) #[1, '<|ca|>']
    forced_decoder_ids.append([2,50262]) #[2, '<|es|>']
    forced_decoder_ids.append([3,50360]) #[3, '<|transcribe|>']

    forced_decoder_ids_modified = forced_decoder_ids
    idx = processor.tokenizer.all_special_tokens.index("<|startofprev|>")
    forced_bos_token_id = processor.tokenizer.all_special_ids[idx]
        
    prompt_tokens = processor.tokenizer(prev_prompt, add_special_tokens=False).input_ids

    # we need to force these tokens
    forced_decoder_ids = []
    for idx, token in enumerate(prompt_tokens):
        # indexing starts from 1 for forced tokens (token at position 0 is the SOS token)
        forced_decoder_ids.append([idx + 1, token])
            
    # now we add the SOS token at the end
    offset = len(forced_decoder_ids)
    forced_decoder_ids.append([offset + 1, model.generation_config.decoder_start_token_id])

    # now we need to append the rest of the prefix tokens (lang, task, timestamps)
    offset = len(forced_decoder_ids)
    for idx, token in forced_decoder_ids_modified:
        forced_decoder_ids.append([idx + offset , token])

    model.generation_config.forced_decoder_ids = forced_decoder_ids

    pred_ids = model.generate(input_features, 
                                    return_timestamps=True,
                                    max_new_tokens=128,
                                    decoder_start_token_id=forced_bos_token_id)
    #exclude prompt from output
    forced_decoder_tokens = convert_forced_to_tokens(forced_decoder_ids)
    output = processor.decode(pred_ids[0][len(forced_decoder_tokens) + 1:], skip_special_tokens=True)
    output_tokens = processor.batch_decode(pred_ids, skip_special_tokens=False)
    return output[1:]

def processing_vad_v3(audio, output_vad, prev_prompt):
    transcription_audio = ""
    first_chunk = True
    for speech in output_vad.get_timeline().support():
        start, end = speech.start, speech.end
        segment_audio = audio[start * 1000:end * 1000]
        segment_audio.export(os.path.join(segments_dir, f"temp_segment.wav"), format="wav")
        filename = os.path.join(segments_dir, f"temp_segment.wav")
        if first_chunk:
            output = generate_1st_chunk(filename)
            first_chunk = False
        else:
            output = generate_from_2nd_chunk(filename, prev_prompt)

        prev_prompt = output
        transcription_audio = transcription_audio + " " + output

    return transcription_audio


def processing_vad_v4(audio, output_vad, threshold, max_duration, prev_prompt, concatenated_segment):
    transcription_audio = ""
    is_first_chunk = True
    for speech in output_vad.get_timeline().support():
        start, end = speech.start, speech.end
        segment_duration = (end - start) * 1000
        segment_audio = audio[start * 1000:end * 1000]

        if max_duration + segment_duration < threshold:
            concatenated_segment += audio[start * 1000:end * 1000]
            max_duration += segment_duration
        else:
            if len(concatenated_segment) > 0:
                temp_segment_path = os.path.join(segments_dir, f"temp_segment.wav")
                concatenated_segment.export(temp_segment_path, format="wav")
                
                if is_first_chunk:
                    output = generate_1st_chunk(temp_segment_path)
                    is_first_chunk = False
                else:
                    output = generate_from_2nd_chunk(temp_segment_path, prev_prompt)

                prev_prompt = output   
                transcription_audio = transcription_audio + output

                max_duration = segment_duration
                concatenated_segment = segment_audio
        
    # Process any remaining audio in the concatenated_segment
    if len(concatenated_segment) > 0:
        temp_segment_path = os.path.join(segments_dir, f"temp_segment.wav")
        concatenated_segment.export(temp_segment_path, format="wav")
        
        output = generate_from_2nd_chunk(temp_segment_path, prev_prompt)
  
        prev_prompt = output
        transcription_audio = transcription_audio + output

    return transcription_audio


def generate(audio_path, use_v4):
    #check audio lenght
    audio = AudioSegment.from_wav(audio_path)
    duration_seconds = len(audio) / 1000.0  

    #apply VAD only if the duration is >30s
    if duration_seconds >= 30:

        output_vad = pipeline_vad(audio_path)
        concatenated_segment = AudioSegment.empty()
        max_duration = 0
        prev_prompt = ""
        if use_v4:
            return processing_vad_v4(audio, output_vad, threshold, max_duration, prev_prompt, concatenated_segment)
        else:
            return  processing_vad_v3(audio, output_vad, prev_prompt)
    else:
        #if duraion is <30s, process directly with generate
        return generate_1st_chunk(audio_path)