textstream-asr 0.1.0

Live speech-to-text streaming on Apple Silicon

TextStream

Live speech-to-text on Apple Silicon. Streams microphone audio through Qwen3-ASR and shows results in a browser — updated every 2.5 seconds with finalized and draft text.

Built for real-time transcription during calls, meetings, or recording sessions. Uses Silero VAD to filter non-speech audio (music, background noise) before it reaches the model, which kills the hallucination problem where ASR models regurgitate their system prompt on noise input.

What it does

• Captures mic audio at 16kHz, runs Silero voice activity detection, feeds speech chunks to Qwen3-ASR on MLX
• Streams finalized + draft text to a browser UI via SSE at localhost:7890
• Saves timestamped transcripts to ~/Documents/textstream/transcripts/YYYY-MM-DD/
• Optionally pushes Grafana annotations for each finalized text segment
• Two model sizes: 0.6B (default, fast) and 1.7B (more accurate), hot-swappable from the browser

Install

pip install textstream-asr

Requires Apple Silicon (M1 or later) and Python 3.10+. MLX doesn't run on Intel or Linux.

Usage

textstream                            # Qwen3-ASR 0.6B, opens browser
textstream --engine qwen-1.7b         # larger model, lower word error rate
textstream --vad-threshold 0.5        # stricter voice detection (default 0.4)
textstream --interval 2.0             # faster updates
textstream --no-browser --no-grafana  # headless
textstream --port 8080                # custom port

The browser UI shows live text with a dark theme. Finalized text in white, draft predictions in grey. Switch models from the dropdown without restarting.

How it works

Every --interval seconds, TextStream drains the mic buffer and runs Silero VAD on the chunk. If speech is detected (probability >= --vad-threshold), the chunk is fed to Qwen3-ASR's streaming decoder. The model returns stable (finalized) text and speculative (draft) text. Stable text is persisted to disk and broadcast to all connected browsers via server-sent events.

If the model hallucinates (outputs its chat system prompt on noise that slips past VAD), a pattern filter catches it and resets the stream. This is a safety net — with VAD active, it almost never fires.

Qwen3-ASR handles its own 30-second sliding context window internally, so there's no manual drift reset needed.

API

GET /          → browser UI
GET /stream    → SSE event stream (data: {"type":"stream","finalized":"...","draft":"..."})
GET /engine    → {"engine":"qwen"} or {"engine":"qwen-1.7b"}
GET /switch?engine=qwen-1.7b → hot-swap model
GET /pause     → pause mic capture
GET /resume    → resume mic capture
GET /stop      → shutdown server

Configuration

Flag	Default	Description
--port	7890	HTTP server port
--engine	qwen	qwen (0.6B) or qwen-1.7b
--interval	2.5	Seconds between transcription updates
--vad-threshold	0.4	Silero VAD speech probability threshold
--no-browser	false	Don't open browser on start
--no-grafana	false	Disable Grafana annotation push

Grafana integration reads GRAFANA_URL and GRAFANA_SERVICE_ACCOUNT_TOKEN from environment variables. If the token is empty, Grafana push is skipped automatically.

Dependencies

• MLX — Apple Silicon ML framework
• mlx-qwen3-asr — Qwen3-ASR for MLX
• silero-vad-lite — Voice activity detection (~2MB, bundles ONNX runtime)
• sounddevice — PortAudio bindings for mic capture
• NumPy

Author

Boris Djordjevic — 199 Biotechnologies

License

MIT