uxoxo/eb2ab

1import os2import numpy as np3import librosa4 5from pyannote.audio import Model6from pyannote.audio.pipelines import VoiceActivityDetection7from lib.conf import tts_dir8from lib.models import default_voice_detection_model9 10class BackgroundDetector:11 12    def __init__(self, wav_file: str):13        self.wav_file   = wav_file14        model = Model.from_pretrained(default_voice_detection_model, cache_dir=tts_dir)15        self.pipeline = VoiceActivityDetection(segmentation=model)16        hyper_params = {17          # onset/offset activation thresholds18          "onset": 0.5, "offset": 0.5,19          # remove speech regions shorter than that many seconds.20          "min_duration_on": 0.0,21          # fill non-speech regions shorter than that many seconds.22          "min_duration_off": 0.023        }24        self.pipeline.instantiate(hyper_params)25 26    def detect(self, vad_ratio_thresh: float=0.05):27        diarization     = self.pipeline(self.wav_file)28        speech_segments = [(s.start, s.end) for s in diarization.get_timeline()]29        total_duration  = librosa.get_duration(path=self.wav_file)30        speech_time     = sum(end - start for start, end in speech_segments)31        non_speech_ratio = 1 - (speech_time / total_duration)32        status = non_speech_ratio > vad_ratio_thresh33        report = {34            'non_speech_ratio': non_speech_ratio,35            'background_detected': status36        }37        return status, report