LiveSRT: Live Speech-to-Text & Translation

LiveSRT is a modular tool for real-time speech-to-text transcription and translation. It captures audio from your microphone (or a file), streams it to state-of-the-art AI transcription providers, and uses Large Language Models (LLMs) to correct and translate the output on the fly, displaying the results in a rich Terminal User Interface (TUI).

📺 Demo

Here's a quick demonstration of LiveSRT in action:

✨ Features

• Live Transcription: Real-time speech-to-text using top-tier providers.
• Live Translation: Translate speech instantly using LLMs (Local or Remote).
• Rich TUI: A dedicated terminal interface to view live transcripts and translations side-by-side.
• Intelligent Post-processing: Uses LLMs to clean up stutters, fix ASR errors, and separate speakers.
• Audio Sources: Support for microphones and audio file replay (via ffmpeg).
• Configurable: Uses a YAML configuration file for reproducible setups.

🔌 Supported Providers

Transcription (ASR)

• AssemblyAI (Streaming API) - Default
• ElevenLabs (Realtime Speech-to-Text)
• Speechmatics (Realtime API)

Translation (LLMs)

• Local LLMs: Runs locally via llama.cpp (e.g., Ministral, Qwen).
• Remote LLMs: Support for Groq, Mistral, Google Gemini, DeepInfra, Ollama and OpenRouter.

  💡 Tip: The best quality/speed model is Ministral 3 8B. We recommend using mistral/ministral-8b-latest, which is free with a Mistral API key.

🚀 Quick Start

As this is a PyPI package, you can run it directly without any installation using uvx (or install via pip).

1. Initialization

First, create a default configuration file. This allows you to select your audio source, transcription backend, and translation settings.

uvx livesrt init-config
# Created config.yml

2. Authentication

Set the API key for your chosen provider (default is AssemblyAI). Keys are stored securely in your system keyring.

uvx livesrt set-token assembly_ai

3. Run

Start the application using the configuration from config.yml.

uvx livesrt run

To enable translation (if disabled in config), you can use the flag:

uvx livesrt run --translate

⚙ Configuration

LiveSRT relies on a config.yml file. You can generate a template using livesrt init-config.

Key Configuration Sections:

• Audio: Select mic or file. If using a microphone, find your device index using livesrt list-microphones.
• Transcription: Choose between assembly_ai, elevenlabs, or speechmatics.
• Translation: Toggle enabled/disabled, choose local-llm or remote-llm, and set source/target languages.
• API Keys: Manage namespaces for multiple environments.

📝 Command Reference

All commands start with livesrt. Use --help on any command for more details.

livesrt init-config

Creates a default config.yml in the current directory.

• --output, -o: Path to the output file (default: config.yml).

livesrt run [OPTIONS]

Runs the main application using the loaded configuration.

• --config, -c: Path to the configuration file (default: config.yml).
• --translate / --no-translate: Override the translation setting in the config.

livesrt set-token <provider> [OPTIONS]

Sets the API token for a specific provider securely.

• <provider> choices:
  • ASR: assembly_ai, elevenlabs, speechmatics
  • LLM: groq, mistral, google, deepinfra, openrouter, ollama
• --api-key, -k: (Optional) Your secret API key. If omitted, you are prompted securely.

livesrt list-microphones

Lists all available input microphone devices and their IDs. Use the resulting ID to update the device_index in your config.yml.

💡 Usage Scenarios

Using a specific microphone

1. List devices: uvx livesrt list-microphones.
2. Edit config.yml: Set audio.device_index to the desired ID.
3. Run: uvx livesrt run.

Debugging with a file

Simulate a live stream using an audio file (requires ffmpeg):

1. Edit config.yml:

   audio:
       source_type: file
       file_path: "./interview.wav"
   

2. Run: uvx livesrt run.

Live Translation with Remote LLM

To offload processing to a fast remote API (e.g., Mistral):

1. Set the key: uvx livesrt set-token mistral.
2. Edit config.yml:

   translation:
       enabled: true
       engine: remote-llm
       remote_llm:
           lang_to: Spanish
           model: mistral/ministral-8b-latest
   

3. Run: uvx livesrt run.

🛠 Development

To set up a local development environment:

uv sync

Development Commands

The Makefile contains helpers for common tasks:

• make format: Formats the code using ruff format.
• make lint: Lints the code using ruff check --fix.
• make types: Performs static type checking using mypy.
• make prettier: Formats Markdown and source files using prettier.
• make clean: Runs all formatters, linters, and type checkers.

🏗 Code Structure

• src/livesrt/cli.py: Entry point and CLI logic using click.
• src/livesrt/containers.py: Dependency Injection container used to wire components based on configuration.
• src/livesrt/tui.py: The Textual-based UI implementation.
• src/livesrt/transcribe/: Audio capture and ASR logic.
  • transcripters/: Implementations for AssemblyAI, ElevenLabs, Speechmatics.
  • audio_sources/: Mic (pyaudio) and File (ffmpeg) sources.
• src/livesrt/translate/: Translation logic.
  • local_llm.py: Wraps llama_cpp for local inference.
  • remote_llm.py: Wraps httpx for OpenAI-compatible APIs.
  • base.py: Handles conversation context and tool-use for accurate translations.

📜 License

This project is licensed under the WTFPL (Do What The Fuck You Want To Public License).