Noah's Local Speech Recognition

This project provides a local, privacy-preserving speech recognition tool using the Vosk speech recognition toolkit and microphone input. It supports keyword detection, timestamped transcription, logging, pause/resume functionality, and model auto-downloading.

Features

• Offline speech recognition (no internet needed)
• Auto-download of Vosk models if not present
• Realtime transcription with timestamps
• Automatic transcript logging to disk
• Support for pause/resume listening

Requirements

• Python 3.9+
• vosk
• pyaudio
• tqdm
• requests

Setup:

pip install noahs_local_speech_recognition

On Linux (including Raspberry Pi), pyaudio still depends on system-level PortAudio. Make sure to install portaudio19-dev and pyaudio is working before pip installing noahs_local_speech_recognition which relies on pyaudio to avoid errors.

Linux Setup Instructions

If you're using Linux, the script attempts to install the required PortAudio libraries automatically. However, you may still need to:

1. Run the script that imports noahs_local_speech_recognition with sudo:

2. Manually install dependencies if issues arise:

   sudo apt-get update
   sudo apt-get install -y portaudio19-dev python3-pyaudio
   pip uninstall pyaudio
   pip install pyaudio
   

These steps ensure that pyaudio is compiled correctly against the system's PortAudio library.

Demo

To test the speech recognizer locally, simply run this script:

from noahs_local_speech_recognition import get_text_after_keyword, list_microphones, start_speech_listening, stop_speech_listening, get_speech_log, get_speech_log_entry, set_speech_log_response, remove_speech_log_entry, pause_speech_listening, resume_speech_listening

import time

print("Listing available microphone devices:")
list_microphones()

print("\nStarting speech recognition with default input device...")
start_speech_listening(name="robot", stop_talking_delay=2, device_index=None, model_name="vosk-model-small-en-us-0.15")

try:
    while True:
        time.sleep(2)
        print("_________________________")
        entry = get_speech_log_entry()
        if entry and entry["response"] is None:
            heard = entry["content"]
            print(f"I HEARD: {heard}")
            speech_log1 = get_speech_log()
            print(f"SPEECH LOG BEFORE SET RESPONSE: {speech_log1}")
            set_speech_log_response("I have heard what you said")
            speech_log2 = get_speech_log()
            print(f"SPEECH LOG AFTER SET RESPONSE: {speech_log2}")
        if entry and entry["content"] in ["goodbye", "good bye", "bye", "quit", "end", "exit"]:
            break
except KeyboardInterrupt:
    pass
finally:
    stop_speech_listening()

This will:

• Auto-download the Vosk "vosk-model-small-en-us-0.15" model if missing.
• Show available audio devices.
• Look for defualt microphone since device_index=None.
• Begin listening and transcribing speech in real-time based on hearing name="robot" keyword
• Log results with timestamps to a file named like transcript_YYYY-MM-DD_HH-MM-SS.txt.
• Display speech_log before and after having response set in the console.
• Use set_speech_log_response() to flag entries as having been responded to.

Changing Models

You can modify the model by altering the model_name passed to start_speech_listening():

start_speech_listening(name="robot", stop_talking_delay=2, device_index=None, model_name="vosk-model-small-en-us-0.15")

Other available models:

• vosk-model-small-en-us-0.15
• vosk-model-en-us-0.22
• vosk-model-en-us-0.22-lgraph

Keyword Mode vs Always Listen Mode

To use Always Listen Mode just use name=None in start_speech_listening():

start_speech_listening(name=None, stop_talking_delay=2, device_index=None, model_name="vosk-model-small-en-us-0.15")

This will transcribe everything, not just things said after the name keyword.

Known Issues

• Use on debian Linux may require manually setting up pyaudio prior to import
• Automaitic Model download links are hardcoded and limited to 3 options.

License

MIT License

Acknowledgements

• Vosk Speech Recognition Toolkit
• TQDM for progress bars