VoxBar AI

How It Works

VoxBar AI uses NVIDIA's Speech-Augmented Language Model (SALM) — a 2.5 billion parameter model that combines a speech encoder with a full Qwen large language model. This means it doesn't just hear sounds and guess words — it understands context, producing transcription with natural punctuation and intelligent word choices.

Here's what happens, step by step:

1. Opens your microphone via sounddevice — captures audio at 16kHz, 1024-sample blocks
2. Buffers 1.5 seconds of audio into a small in-memory buffer (~96KB)
3. Checks for silence — if the RMS energy is below 0.03, the chunk is discarded (no wasted processing)
4. Writes a tiny temp WAV file (~48KB) to your system temp folder
5. Feeds the WAV to the SALM model with a structured chat prompt: "Transcribe this audio to English"
6. The model generates text using its LLM backbone — producing accurate, contextual transcription
7. Token output is decoded via the model's tokenizer into readable text
8. Temp file is immediately deleted — nothing accumulates on disk
9. Text is appended to your textbox
10. Repeats forever — each chunk is completely independent

Every chunk is a self-contained operation. There's no growing context window, no accumulating state, no connection to maintain. The GPU just processes one tiny WAV file at a time, indefinitely.

Recording Limits

VoxBar AI Has No Recording Limit

VoxBar AI can record continuously for as long as you want — hours, all day if you need it.

Unlike VoxBar Pro, which runs inside a Docker container with WebSocket connections that can drop, VoxBar AI runs natively on your machine. There is no server process, no container, no network connection involved. It's just your Python process, the model in GPU memory, and your microphone.

Why It Runs Forever

• Each 1.5-second chunk is completely independent — no state carries over
• GPU memory is fixed — the same model processes the same size input every time
• No WebSocket, no Docker, no server to crash or restart
• No context window that fills up or degrades

Auto-Stop Behaviour

• Silence timeout: 15 minutes (900 seconds) of no detected speech
• Check interval: Every 10 seconds
• This means you can pause for a long coffee break, step away from your desk, or sit in a quiet meeting — VoxBar AI will keep waiting patiently for up to 15 minutes before auto-stopping

Real-World Testing (2026-02-17)

During live testing, VoxBar AI ran continuously for over 50 minutes of natural dictation with zero interruptions, zero restarts, and zero degradation. The text quality at minute 50 was identical to minute 1.

Memory & Resource Footprint

If you left VoxBar AI running for a 3-hour meeting, it would just keep transcribing every 1.5 seconds without ever stopping, restarting, or degrading. That's a genuine advantage over container-based solutions.

Architecture Advantage

What makes VoxBar AI special: It combines a FastConformer speech encoder with a Qwen 2.5B large language model. This is fundamentally different from traditional ASR models that just map sounds to words. The LLM backbone means:

• Context-aware transcription — it understands what you're saying, not just what sounds you make
• Natural punctuation — periods, commas, and question marks appear where they should
• Intelligent word choices — homophones and ambiguous sounds are resolved using language understanding
• No hallucination on silence — the silence filter (0.03 RMS threshold) prevents phantom text

What users DON'T have to worry about:
- ❌ No Docker required — runs natively, no containers
- ❌ No internet connection — completely offline
- ❌ No WebSocket connections — nothing to drop or reconnect
- ❌ No session limits — record for hours without interruption
- ❌ No cloud processing — your voice never leaves your machine
- ❌ No API keys — the model runs locally
- ❌ No usage limits — unlimited transcription, forever

What users DO need to know:
- ⚠️ Text arrives in chunks (every ~1.5 seconds), not word-by-word like VoxBar Pro
- ⚠️ NVIDIA GPU required — the SALM model needs CUDA
- ⚠️ 6-8GB VRAM — needs a mid-range or better NVIDIA GPU
- ⚠️ First launch downloads ~5GB model files (cached after that)

Accuracy & Speed

Real-World Accuracy

During live testing, VoxBar AI captured natural dictation with zero editing required. Every word was captured accurately, including technical terms, proper nouns, and conversational speech. The accuracy is functionally identical to VoxBar Pro (Voxtral) — the only difference is delivery speed.

Hardware Requirements

License & Attribution

Compared to VoxBar Pro

Bottom line: VoxBar AI delivers the same accuracy as the flagship, with none of the Docker complexity, and it never stops. For users who don't need word-by-word real-time display, VoxBar AI is arguably the more practical choice.