Skip to content

Latest commit

 

History

History
44 lines (35 loc) · 2.05 KB

README_en.md

File metadata and controls

44 lines (35 loc) · 2.05 KB
Raw

SpeakEasy

中文版本

Welcome to SpeakEasy, a simple yet useful project that combines off-the-shelf LLM models, automatic speech recognition (ASR), and text-to-speech (TTS) methods to create an interactive spoken chatbot.

With SpeakEasy, you can enhance your oral English skills by engaging in conversations with the LLM model. It operates in real-time and offline on your laptop or PC.

Tested on M2 16G MacBook Air / (4070Ti) + 13600KF Ubuntu / (4070Ti) + 13600KF Windows. demo

Features

  • Practice oral English through interactive conversations
  • Real-time and offline functionality
  • Utilizes ChatGLM3 6B 4-bit quantized model for chat interactions
  • Accelerated automatic speech recognition (ASR) with whisper.cpp
  • Text-to-speech (TTS) conversion using EfficientSpeech

Usage

  1. Follow the README of Chatglm.cpp to install chatglm.cpp
  2. Download ChatGLM3 6B-4bit model
  3. Install whisper.cpp and compile with BLAS / CUBLAS can speed up the inference process
  4. Install EfficientSpeech for real-time TTS
  5. Install requirements pip install -r requirements.txt
  6. Create local TTS service cd examples/efficientspeech/ && sh es_tts_service.sh
  7. Modify the model path of the script and run it: 
    vim examples/demo.sh 
cd examples && sh demo.sh

Tips

  • Press 'SPACE' key to start recording your voice, and press it again to finish.
  • If the input device is not your build-in device, modify the --input-device to your real input device index. Find the right device index via: 
    import sounddevice as sd
input_devices = sd.query_devices(kind='input')

Credits