Python 2.7（Windows 10环境）说话人识别技术求助

Hey there! Since you're on Windows 10 using Python 2.7 and the speech_recognition library, and you want to build a system that only recognizes your voice and what you say, let's break this down into manageable steps—perfect for a beginner!

Core Idea: Combine Speech-to-Text + Speaker Verification #

Regular speech recognition only converts audio to text, so you need an extra layer: speaker verification (checking if the voice matches yours). Here's how to do this with tools compatible with Python 2.7:

1. Recommended Libraries for Speaker Verification #

These libraries work with Python 2.7 and can handle voice identification:

• pyAudioAnalysis: A versatile library that includes speaker recognition features. It can extract unique voice features from your samples and compare them to incoming audio.
  • Install it via pip2: pip2 install pyaudioanalysis
  • You'll also need pyaudio for recording: pip2 install pyaudio
• CMU PocketSphinx: While mainly for speech-to-text, it supports speaker identification too. It integrates well with the speech_recognition library you're already using.
  • Install: pip2 install pocketsphinx

2. Step-by-Step Implementation (Using pyAudioAnalysis) #

Let's walk through a basic workflow:

1. Record Your Voice Samples: Record 5-10 short audio clips (10-30 seconds each) of you speaking in different tones/backgrounds, save them as WAV files in a folder (e.g., my_voice_samples).
2. Train a Speaker Model: Run this code once to create your unique voice template:

   from pyAudioAnalysis import audioTrainTest as aT
   # Train an SVM model with your samples
   aT.featureAndTrain(["my_voice_samples"], 1.0, 1.0, 0.05, 0.02, "svm", "my_speaker_model", False)
   

3. Real-Time Recognition + Verification: Combine this with speech_recognition to listen, verify, and transcribe only your voice:

   import speech_recognition as sr
   import os
   from pyAudioAnalysis import audioTrainTest as aT
   
   # Initialize the speech recognizer
   recognizer = sr.Recognizer()
   
   with sr.Microphone() as source:
       print("Speak now...")
       # Capture audio from the mic
       audio = recognizer.listen(source)
   
       # Save temporary WAV file for speaker verification
       temp_audio_path = "temp_voice.wav"
       with open(temp_audio_path, "wb") as f:
           f.write(audio.get_wav_data())
   
       # Check if the voice matches yours
       _, prediction_results, _ = aT.fileClassification(temp_audio_path, "my_speaker_model", "svm")
       # Set a threshold (adjust based on your tests)
       confidence_threshold = 0.8
       if prediction_results[0] > confidence_threshold:
           print("Recognized your voice! Transcribing...")
           try:
               # Convert audio to text (using Google's API)
               text = recognizer.recognize_google(audio, language="zh-CN")
               print(f"You said: {text}")
           except sr.UnknownValueError:
               print("Sorry, I couldn't understand what you said.")
           except sr.RequestError as e:
               print(f"Couldn't connect to the speech service: {e}")
       else:
           print("This isn't your voice—ignoring input.")
   
       # Clean up temporary file
       os.remove(temp_audio_path)
   

3. Pro Tips for Better Accuracy #

• Record Samples in Different Scenarios: Include clips with slight background noise, different volumes, and different phrases to make your model more robust.
• Adjust the Confidence Threshold: If the system is too strict (rejects your voice often), lower the threshold a bit; if it accepts others too easily, raise it.
• Note on Python 2.7: Python 2.7 is no longer supported, so if you can, consider upgrading to Python 3.x later—you'll get access to more modern speaker recognition libraries like Resemblyzer or SpeechBrain.

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