基于Python的说话人识别工具问询：如何用已有数据识别录音说话人

当然存在适用于Python的类库，能基于已有语音数据完成说话人身份识别，下面是几个实用选项和贴近你需求的代码实现：

常用类库 #

• pyannote.audio：基于深度学习的语音处理工具，自带预训练模型，可快速实现说话人匹配
• SpeechBrain：开源语音AI框架，提供模块化的说话人验证/识别组件，灵活性强
• librosa + scikit-learn：手动提取语音特征（如梅尔频谱）+ 传统机器学习分类，适合自定义实现

实现示例（基于pyannote.audio） #

这个方案贴近你的示例逻辑，用预训练模型快速完成识别：

首先安装依赖：

pip install pyannote.audio sounddevice soundfile scipy

然后编写代码：

import sounddevice as sd
import soundfile as sf
from pyannote.audio import Pipeline
from scipy.spatial.distance import cosine

# 替换为你在Hugging Face获取的免费访问令牌
ACCESS_TOKEN = "你的Hugging Face令牌"

# 初始化说话人识别管道
pipeline = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    use_auth_token=ACCESS_TOKEN
)

# 预存的目标说话人语音文件
target_voices = ['obama.wav', 'trump.wav', 'biden.wav']
# 预加载并提取每个说话人的特征嵌入
speaker_embeddings = {}
for voice_path in target_voices:
    # 处理单说话人音频，提取嵌入
    diarization = pipeline(voice_path)
    speaker_segment = next(diarization.itersegments())
    speaker_embeddings[voice_path.split('.')[0]] = speaker_segment[2]

def record_audio(output_file='recorded.wav', duration=5, samplerate=16000):
    """录制语音并保存为WAV文件"""
    print("录音中...")
    audio_data = sd.rec(int(duration * samplerate), samplerate=samplerate, channels=1)
    sd.wait()
    sf.write(output_file, audio_data, samplerate)
    return output_file

# 录制待识别语音
recorded_file = record_audio()
# 提取录制语音的说话人嵌入
recorded_diarization = pipeline(recorded_file)
recorded_embedding = next(recorded_diarization.itersegments())[2]

# 匹配最相似的说话人
min_dist = float('inf')
matched_speaker = None
for speaker, embedding in speaker_embeddings.items():
    dist = cosine(recorded_embedding, embedding)
    if dist < min_dist:
        min_dist = dist
        matched_speaker = speaker

print(f"The speaker is likely {matched_speaker}")

轻量实现（基于librosa + scikit-learn） #

如果不想依赖深度学习模型，可以用传统机器学习方案：

安装依赖：

pip install librosa scikit-learn sounddevice soundfile

代码实现：

import librosa
import numpy as np
import sounddevice as sd
import soundfile as sf
from sklearn.svm import SVC
from sklearn.preprocessing import StandardScaler

def extract_mfcc_features(file_path):
    """提取语音的MFCC特征（均值）"""
    y, sr = librosa.load(file_path, sr=16000)
    mfcc_features = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
    return np.mean(mfcc_features, axis=1)

def record_audio(output_file='recorded.wav', duration=5, samplerate=16000):
    print("录音中...")
    audio_data = sd.rec(int(duration * samplerate), samplerate=samplerate, channels=1)
    sd.wait()
    sf.write(output_file, audio_data, samplerate)
    return output_file

# 预存目标语音
target_voices = ['obama.wav', 'trump.wav', 'biden.wav']
# 提取特征并准备训练数据
X = []
labels = []
for voice_path in target_voices:
    feat = extract_mfcc_features(voice_path)
    X.append(feat)
    labels.append(voice_path.split('.')[0])

# 训练分类器
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
classifier = SVC(kernel='linear')
classifier.fit(X_scaled, labels)

# 录制并识别
recorded_file = record_audio()
recorded_feat = extract_mfcc_features(recorded_file)
recorded_feat_scaled = scaler.transform([recorded_feat])
matched_speaker = classifier.predict(recorded_feat_scaled)[0]

print(f"The speaker is likely {matched_speaker}")

内容的提问来源于stack exchange，提问作者plshelpmeout