diarization.py

import os
import torch
import math
from moviepy.editor import VideoFileClip, AudioFileClip
from pyannote.audio import Pipeline
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import librosa
import datetime
from collections import defaultdict
import numpy as np
import spaces

class LazyDiarizationPipeline:
    def __init__(self):
        self.pipeline = None

    @spaces.GPU(duration=120)
    def get_pipeline(self, diarization_access_token):
        if self.pipeline is None:
            self.pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", use_auth_token=diarization_access_token)
            self.pipeline = self.pipeline.to(torch.device("cuda"))
        return self.pipeline

lazy_diarization_pipeline = LazyDiarizationPipeline()

class LazyTranscriptionPipeline:
    def __init__(self):
        self.model = None
        self.processor = None
        self.pipe = None

    @spaces.GPU(duration=120)
    def get_pipeline(self, language):
        if self.pipe is None:
            model_id = "openai/whisper-large-v3"
            self.model = AutoModelForSpeechSeq2Seq.from_pretrained(
                model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True, use_safetensors=True
            )
            self.model.to(torch.device("cuda"))
            self.processor = AutoProcessor.from_pretrained(model_id)
            self.pipe = pipeline(
                "automatic-speech-recognition",
                model=self.model,
                tokenizer=self.processor.tokenizer,
                feature_extractor=self.processor.feature_extractor,
                max_new_tokens=128,
                chunk_length_s=30,
                batch_size=1,
                return_timestamps=True,
                torch_dtype=torch.float16,
                device=torch.device("cuda"),
                generate_kwargs={"language": language}
            )
        return self.pipe

lazy_transcription_pipeline = LazyTranscriptionPipeline()

def extract_audio(video_path, audio_path):
    video = VideoFileClip(video_path)
    audio = video.audio
    audio.write_audiofile(audio_path, codec='pcm_s16le', fps=16000)

def format_timestamp(seconds):
    return str(datetime.timedelta(seconds=seconds)).split('.')[0]

@spaces.GPU(duration=100)
def transcribe_audio(audio_path, language):
    pipe = lazy_transcription_pipeline.get_pipeline(language)

    audio, sr = librosa.load(audio_path, sr=16000)
    duration = len(audio) / sr
    n_chunks = math.ceil(duration / 30)
    transcription_txt = ""
    transcription_chunks = []

    for i in range(n_chunks):
        start = i * 30 * sr
        end = min((i + 1) * 30 * sr, len(audio))
        audio_chunk = audio[start:end]

        # Convert the audio chunk to float32 numpy array
        audio_chunk = (audio_chunk * 32767).astype(np.float32)

        result = pipe(audio_chunk)
        transcription_txt += result["text"]
        for chunk in result["chunks"]:
            start_time, end_time = chunk["timestamp"]
            transcription_chunks.append({
                "start": start_time + i * 30,
                "end": end_time + i * 30,
                "text": chunk["text"]
            })

        print(f"Transcription Progress: {int(((i + 1) / n_chunks) * 100)}%")

    return transcription_txt, transcription_chunks

def create_combined_srt(transcription_chunks, diarization, output_path):
    speaker_segments = []
    speaker_map = {}
    current_speaker_num = 1

    for segment, _, speaker in diarization.itertracks(yield_label=True):
        if speaker not in speaker_map:
            speaker_map[speaker] = f"Speaker {current_speaker_num}"
            current_speaker_num += 1
        speaker_segments.append((segment.start, segment.end, speaker_map[speaker]))

    with open(output_path, 'w', encoding='utf-8') as srt_file:
        for i, chunk in enumerate(transcription_chunks, 1):
            start_time, end_time = chunk["start"], chunk["end"]
            text = chunk["text"]

            # Find the corresponding speaker
            current_speaker = "Unknown"
            for seg_start, seg_end, speaker in speaker_segments:
                if seg_start <= start_time < seg_end:
                    current_speaker = speaker
                    break

            # Format timecodes as h:mm:ss (without leading zeros for hours)
            start_str = format_timestamp(start_time).split('.')[0].lstrip('0')
            end_str = format_timestamp(end_time).split('.')[0].lstrip('0')

            srt_file.write(f"{i}\n")
            srt_file.write(f"{{{current_speaker}}}\n time: ({start_str} --> {end_str})\n text: {text}\n\n")

    # Add dominant speaker information
    speaker_durations = defaultdict(float)
    for seg_start, seg_end, speaker in speaker_segments:
        speaker_durations[speaker] += seg_end - seg_start

    dominant_speaker = max(speaker_durations, key=speaker_durations.get)
    dominant_duration = speaker_durations[dominant_speaker]

    with open(output_path, 'a', encoding='utf-8') as srt_file:
        dominant_duration_str = format_timestamp(dominant_duration).split('.')[0].lstrip('0')
        srt_file.write(f"\nMost dominant speaker: {dominant_speaker} with total duration {dominant_duration_str}\n")

@spaces.GPU(duration=100)
def process_video(video_path, diarization_access_token, language):
    base_name = os.path.splitext(video_path)[0]
    audio_path = f"{base_name}.wav"
    extract_audio(video_path, audio_path)

    # Diarization
    print("Performing diarization...")
    pipeline = lazy_diarization_pipeline.get_pipeline(diarization_access_token)
    diarization = pipeline(audio_path)
    print("Diarization complete.")

    # Transcription
    print("Performing transcription...")
    transcription, chunks = transcribe_audio(audio_path, language)
    print("Transcription complete.")

    # Create combined SRT file
    combined_srt_path = f"{base_name}_combined.srt"
    create_combined_srt(chunks, diarization, combined_srt_path)
    print(f"Combined SRT file created and saved to {combined_srt_path}")

    # Clean up
    os.remove(audio_path)

    return combined_srt_path