SpeechToText

Overview

This script implements a speech recognition service that listens for audio input and provides recognized text through a FastAPI-based web server. It uses a speech recognition engine (default: Vosk) to continuously process speech from the microphone and updates the recognized text in real time.

The FastAPI server offers multiple endpoints for connectivity, including WebSocket support for real-time updates to connected clients.

The script runs the speech recognition loop in a separate thread, allowing it to continuously capture speech while the server handles API requests and WebSocket connections concurrently.

Instructions

Warning: Do not use with Python 3.13, as most plugins are not yet compatible. Instead, use Python 3.11 or 3.12.
I personally use 3.12.8 at the moment.

How to run

... if you have multiple python installations (and the default installation is not 3.12):

• In your terminal locate the Python312 folder.
  For me it's: C:\Users\Admin\AppData\Local\Programs\Python\Python312

Make sure the terminal runs on this folder!

• In the terminal run: python.exe -m pip install RequiredPackageNameHere
  Refer to the package section below.

• In VSCode or your editor of choice open a terminal (doesn't need to be on the Python312 folder path.)
  and run: py -V:3.12 .\speech_to_text.py

OR (Recommended) you can setup a venv (virtual environment) based on your python 3.12 installation.

OR you can switch the global installation by going into the windows enviroment variables and changing the entries

• C:\Users\Admin\AppData\Local\Programs\Python\Python313\Scripts\

and

• C:\Users\Admin\AppData\Local\Programs\Python\Python313\

in the system (variable) path to use Python312 (make sure python 3.12 is installed).

---

When you want to target a different Vosk model for speech recognition just change the vosk_model_path path in the corresponding config.json (example: keyword_recognition_config.json).

You can download additional Vosk models from https://alphacephei.com/vosk/models.

Required Packages

Important:
Check the requirements.txt for a detailed overview!

• PyAudio
  PyAudio provides Python bindings for PortAudio v19, the cross-platform audio I/O library.

• soundfile
  Library for reading and writing sound files.

• uvicorn
  Uvicorn is an ASGI web server implementation for Python.

• fastapi
  FastAPI is a modern, high-performance, web framework for building APIs with Python based on standard Python type hints.

• Levenshtein
  Levenshtein is a string metric that measures the minimum number of single-character edits (insertions, deletions, etc.). In STT the package is useful for implementing fuzzy matching to account for mispronunciations or transcription errors in keyword detection.

---

Depending on the model(s) you want to use, not all of these packages may be required. Select the ones you need.

• vosk (Recommended!)
  Vosk is an offline open source speech recognition toolkit.

• tensorflow
  TensorFlow is an open source software library for high performance numerical computation.

• openai-whisper

  • torch
    PyTorch, a deep learning framework.

  • whisper (Not sure if required!)
    A text-to-speech library.

  • openai-whisper
    OpenAI's Whisper model for speech-to-text transcription.

Sources

Speech recognition in python (great for general learning)

• https://realpython.com/python-speech-recognition/
• https://www.simplilearn.com/tutorials/python-tutorial/speech-recognition-in-python
• https://medium.com/naukri-engineering/speech-recognition-in-python-fcda027a97a1