| import webrtcvad | |
| import numpy as np | |
| import librosa | |
| def apply_vad(audio, sr, frame_duration=30, aggressiveness=3): | |
| ''' | |
| Voice Activity Detection (VAD): It is a technique used to determine whether a segment of audio contains speech. | |
| This is useful in noisy environments where you want to filter out non-speech parts of the audio. | |
| webrtcvad: This is a Python package based on the VAD from the WebRTC (Web Real-Time Communication) project. | |
| It helps detect speech in small chunks of audio. | |
| ''' | |
| vad = webrtcvad.Vad() | |
| audio_int16 = np.int16(audio * 32767) | |
| frame_size = int(sr * frame_duration / 1000) | |
| frames = [audio_int16[i:i + frame_size] for i in range(0, len(audio_int16), frame_size)] | |
| voiced_audio = np.concatenate([frame for frame in frames if vad.is_speech(frame.tobytes(), sample_rate=sr)]) | |
| voiced_audio = np.float32(voiced_audio) / 32767 | |
| return voiced_audio | |
| # import webrtcvad | |
| # import librosa | |
| # import numpy as np | |
| # def apply_vad(audio, sr, frame_duration_ms=30): | |
| # # Initialize WebRTC VAD | |
| # vad = webrtcvad.Vad() | |
| # vad.set_mode(1) # Set aggressiveness mode (0-3) | |
| # # Convert to 16kHz if not already | |
| # if sr != 16000: | |
| # audio = librosa.resample(audio, orig_sr=sr, target_sr=16000) | |
| # sr = 16000 | |
| # # Convert to 16-bit PCM | |
| # audio = (audio * 32768).astype(np.int16) | |
| # frame_length = int(sr * (frame_duration_ms / 1000.0)) # Calculate fram | |
| # e length in samples | |
| # bytes_per_frame = frame_length * 2 # 16-bit audio has 2 bytes per sample | |
| # # Apply VAD to the audio | |
| # voiced_frames = [] | |
| # for i in range(0, len(audio), frame_length): | |
| # frame = audio[i:i + frame_length].tobytes() | |
| # if len(frame) == bytes_per_frame and vad.is_speech(frame, sr): | |
| # voiced_frames.extend(audio[i:i + frame_length]) | |
| # # Return the VAD-filtered audio | |
| # return np.array(voiced_frames) | |
| # import webrtcvad | |
| # import numpy as np | |
| # import librosa | |
| # def apply_vad(audio, sr, frame_duration=30, aggressiveness=3): | |
| # ''' | |
| # Voice Activity Detection (VAD): Detects speech in audio. | |
| # ''' | |
| # vad = webrtcvad.Vad(aggressiveness) | |
| # # Resample to 16000 Hz if not already (recommended for better compatibility) | |
| # if sr != 16000: | |
| # audio = librosa.resample(audio, orig_sr=sr, target_sr=16000) | |
| # sr = 16000 | |
| # # Convert to 16-bit PCM format expected by webrtcvad | |
| # audio_int16 = np.int16(audio * 32767) | |
| # # Ensure frame size matches WebRTC's expected lengths | |
| # frame_size = int(sr * frame_duration / 1000) | |
| # if frame_size % 2 != 0: | |
| # frame_size -= 1 # Make sure it's even to avoid processing issues | |
| # frames = [audio_int16[i:i + frame_size] for i in range(0, len(audio_int16), frame_size)] | |
| # # Filter out non-speech frames | |
| # voiced_frames = [] | |
| # for frame in frames: | |
| # if len(frame) == frame_size and vad.is_speech(frame.tobytes(), sample_rate=sr): | |
| # voiced_frames.append(frame) | |
| # # Concatenate the voiced frames | |
| # voiced_audio = np.concatenate(voiced_frames) | |
| # voiced_audio = np.float32(voiced_audio) / 32767 | |
| # return voiced_audio | |