Spaces:
Runtime error
Runtime error
File size: 5,332 Bytes
9deffb0 a08b017 4ee6f09 9deffb0 a08b017 156337d a08b017 156337d a08b017 156337d a08b017 156337d a08b017 0d98195 a08b017 0d98195 a08b017 0d98195 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 |
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,
chunk_length_s=30,
return_timestamps=True,
device=torch.device("cuda")
)
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]
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"]
})
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"]
current_speaker = "Unknown"
for seg_start, seg_end, speaker in speaker_segments:
if seg_start <= start_time < seg_end:
current_speaker = speaker
break
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"{start_str} --> {end_str}\n")
srt_file.write(f"{current_speaker}: {text}\n\n")
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)
pipeline = lazy_diarization_pipeline.get_pipeline(diarization_access_token)
diarization = pipeline(audio_path)
transcription, chunks = transcribe_audio(audio_path, language)
combined_srt_path = f"{base_name}_combined.srt"
create_combined_srt(chunks, diarization, combined_srt_path)
os.remove(audio_path)
return combined_srt_path
|