import streamlit as st import streamlit.components.v1 as stc import noisereduce as nr import librosa import soundfile as sf import numpy as np import plotly.graph_objects as go import pickle from pyannote.audio.utils.signal import Binarize import torch @st.cache def speech_activity_detection_model(): # sad = torch.hub.load('pyannote-audio', 'sad_ami', source='local', device='cpu', batch_size=128) with open('speech_activity_detection_model.pkl', 'rb') as f: sad = pickle.load(f) return sad @st.cache def trim_noise_part_from_speech(sad, fname, speech_wav, sr): file_obj = {"uri": "filename", "audio": fname} sad_scores = sad(file_obj) binarize = Binarize(offset=0.52, onset=0.52, log_scale=True, min_duration_off=0.1, min_duration_on=0.1) speech = binarize.apply(sad_scores, dimension=1) noise_wav = np.zeros((speech_wav.shape[0], 0)) append_axis = 1 if speech_wav.ndim == 2 else 0 noise_ranges = [] noise_start = 0 for segmentation in speech.segmentation(): noise_end, next_noise_start = int(segmentation.start*sr), int(segmentation.end*sr) noise_wav = np.append(noise_wav, speech_wav[:, noise_start:noise_end], axis=append_axis) noise_ranges.append((noise_start/sr, noise_end/sr)) noise_start = next_noise_start return noise_wav.T, noise_ranges @st.cache def trim_audio(data, rate, start_sec=None, end_sec=None): start, end = int(start_sec * rate), int(end_sec * rate) if data.ndim == 1: # mono return data[start:end] elif data.ndim == 2: # stereo return data[:, start:end] title = 'Audio noise reduction' st.set_page_config(page_title=title, page_icon=":sound:") st.title(title) uploaded_file = st.file_uploader("Upload your audio file (.wav)") is_file_uploaded = uploaded_file is not None if not is_file_uploaded: uploaded_file = 'sample.wav' wav, sr = librosa.load(uploaded_file, sr=None) wav_seconds = int(len(wav)/sr) st.subheader('Original audio') st.audio(uploaded_file) st.subheader('Noise part') noise_part_detection_method = st.radio('Noise source detection', ['Manually', 'Automatically (using speech activity detections)']) if noise_part_detection_method == "Manually": # ノイズ区間は1箇所 default_ranges = (0.0, float(wav_seconds)) if is_file_uploaded else (73.0, float(wav_seconds)) noise_part_ranges = [st.slider("Select a part of the noise (sec)", 0.0, float(wav_seconds), default_ranges, step=0.1)] noise_wav = trim_audio(wav, sr, noise_part_ranges[0][0], noise_part_ranges[0][1]) elif noise_part_detection_method == "Automatically (using speech activity detections)": # ノイズ区間が複数 with st.spinner('Please wait for Detecting the speech activities'): sad = speech_activity_detection_model() noise_wav, noise_part_ranges = trim_noise_part_from_speech(sad, uploaded_file, wav, sr) fig = go.Figure() x_wav = np.arange(len(wav)) / sr fig.add_trace(go.Scatter(y=wav[::1000])) for noise_part_range in noise_part_ranges: fig.add_vrect(x0=int(noise_part_range[0]*sr/1000), x1=int(noise_part_range[1]*sr/1000), fillcolor="Red", opacity=0.2) fig.update_layout(width=700, margin=dict(l=0, r=0, t=0, b=0, pad=0)) fig.update_yaxes(visible=False, ticklabelposition='inside', tickwidth=0) st.plotly_chart(fig, use_container_with=True) st.text('Noise audio') sf.write('noise_clip.wav', noise_wav, sr) noise_wav, sr = librosa.load('noise_clip.wav', sr=None) st.audio('noise_clip.wav') if st.button('Denoise the audio!'): with st.spinner('Please wait for completion'): nr_wav = nr.reduce_noise(audio_clip=wav, noise_clip=noise_wav, prop_decrease=1.0) st.subheader('Denoised audio') sf.write('nr_clip.wav', nr_wav, sr) st.success('Done!') st.text('Denoised audio') st.audio('nr_clip.wav')