pip install pyannote.audio
场景：
一段音频中有多个说话人，将不同的人说的话分离出来
已知一些人的语音特征，跟分离出来的片段，分别求特征的余弦距离，余弦距离最小的作为说话的人
_*_ coding: utf-8 _*_
# @Time : 2024/3/16 10:47
# @Author : Michael
# @File : spearker_rec.py
# @desc :
import torch
from pyannote.audio import Model, Pipeline, Inference
from pyannote.core import Segment
from scipy.spatial.distance import cosine
def extract_speaker_embedding(pipeline, audio_file, speaker_label):
diarization = pipeline(audio_file)
speaker_embedding = None
for turn, _, label in diarization.itertracks(yield_label=True):
if label == speaker_label:
segment = Segment(turn.start, turn.end)
speaker_embedding = inference.crop(audio_file, segment)
break
return speaker_embedding
# 对于给定的音频，提取声纹特征并与人库中的声纹进行比较
def recognize_speaker(pipeline, audio_file):
diarization = pipeline(audio_file)
speaker_turns = []
for turn, _, speaker_label in diarization.itertracks(yield_label=True):
# 提取切片的声纹特征
embedding = inference.crop(audio_file, turn)
distances = {}
for speaker, embeddings in speaker_embeddings.items():
# 计算与已知说话人的声纹特征的余弦距离
distances[speaker] = min([cosine(embedding, e) for e in embeddings])
# 选择距离最小的说话人
recognized_speaker = min(distances, key=distances.get)
speaker_turns.append((turn, recognized_speaker))
# 记录说话人的时间段和余弦距离最小的预测说话人
return speaker_turns
if __name__ == "__main__":
token = "hf_***" # 请替换为您的Hugging Face Token
# 加载声音分离识别模型
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-3.1",
use_auth_token=token, # 在项目页面agree使用协议，并获取 Hugging Face Token
# cache_dir="/home/huggingface/hub/models--pyannote--speaker-diarization-3.1/"
)
# 加载声纹嵌入模型
embed_model = Model.from_pretrained("pyannote/embedding", use_auth_token=token)
inference = Inference(embed_model, window="whole")
# pipeline.to(torch.device("cuda"))
# 假设您已经有一个包含不同人声的音频文件集,以及对应的人
audio_files = {
"mick": "mick.wav", # mick的音频
"moon": "moon.wav", # moon的音频
}
speaker_embeddings = {}
for speaker, audio_file in audio_files.items():
diarization = pipeline(audio_file)
for turn, _, speaker_label in diarization.itertracks(yield_label=True):
embedding = extract_speaker_embedding(pipeline, audio_file, speaker_label)
# 获取原始已知说话人的声纹特征
speaker_embeddings.setdefault(speaker, []).append(embedding)
# 给定新的未知人物的音频文件
given_audio_file = "2_voice.wav" # 前半部分是 mick 说话，后半部分是 moon 说话
# 识别给定音频中的说话人
recognized_speakers = recognize_speaker(pipeline, given_audio_file)
print("Recognized speakers in the given audio:")
for turn, speaker in recognized_speakers:
print(f"Speaker {speaker} spoke between {turn.start:.2f}s and {turn.end:.2f}s")
输出：
Model was trained with pyannote.audio 0.0.1, yours is 3.1.1. Bad things might happen unless you revert pyannote.audio to 0.x.
Model was trained with torch 1.8.1+cu102, yours is 2.2.1+cpu. Bad things might happen unless you revert torch to 1.x.
Recognized speakers in the given audio:
Speaker mick spoke between 0.57s and 1.67s
Speaker moon spoke between 2.47s and 2.81s
Speaker moon spoke between 3.08s and 4.47s