VoxBar AI: Inside NVIDIA Canary Qwen 2.5B
The FastConformer + Qwen SALM architecture that gives VoxBar AI its context awareness, natural punctuation, and multilingual support — without needing Docker.
The Model: NVIDIA Canary Qwen 2.5B
NVIDIA's Canary Qwen 2.5B is a Speech-Augmented Language Model (SALM) built on the NeMo framework. It pairs NVIDIA's FastConformer speech encoder with Alibaba's Qwen language model backbone, creating an architecture that doesn't just hear sounds — it understands context. Each audio chunk is processed independently, producing clean text with natural punctuation, capitalisation, and correct word choices.
This is fundamentally different from VoxBar Pro's Voxtral model, which streams token-by-token. The trade-off: Canary has slightly higher latency per chunk, but each chunk is fully context-aware, producing clean text with built-in punctuation and capitalisation.
The Architecture
FastConformer Encoder (24 layers)
The encoder uses NVIDIA's FastConformer — a speed-optimised variant of the Conformer architecture. Conformers combine convolutional neural networks (for local audio features) with self-attention (for global context). FastConformer adds multi-head attention optimisations and downsampling improvements that make it significantly faster at inference time without sacrificing accuracy.
The encoder processes 16kHz mono audio and outputs a sequence of audio feature vectors — a compressed representation of the speech signal.
Qwen Language Backbone
Instead of a standard Transformer decoder, Canary Qwen 2.5B uses Alibaba's Qwen language model as its decoder. This gives it genuine language understanding — the ability to distinguish "there/their/they're" from context, produce natural sentence structure, and handle domain-specific vocabulary. Task tokens at the start of the sequence control the output language and formatting.
Why "Runs Indefinitely"?
Because Canary is chunk-based, each audio segment is processed independently. There's no growing context window, no accumulated state, no memory leak. The model uses the exact same VRAM on minute 1 as on minute 60. This architecture is inherently long-session stable — you can transcribe all day without any degradation.
This contrasts with streaming models (like VoxBar Pro's Voxtral) which need session refresh logic to avoid quality drift over time.
No Docker Required
VoxBar AI runs as a native Python application. The Canary Qwen 2.5B model is loaded directly through PyTorch and NVIDIA's NeMo toolkit — no containers, no Docker Desktop, no virtualisation overhead. You launch the app, it downloads the model on first run, and you start talking.
This makes VoxBar AI the most accessible premium engine in the VoxBar lineup. If you have an NVIDIA GPU with 6GB+ VRAM, it just works.
System Audio Capture
Like all VoxBar tiers, VoxBar AI supports system audio capture — transcribing your PC's audio output directly. Meetings, podcasts, YouTube tutorials — anything playing on your machine can be transcribed locally, without cloud upload.
Smart Autocorrect & Mid-Text Editing
Between chunks, VoxBar AI's autocorrect cleans up double spaces, fixes capitalisation, and normalises formatting. Combined with mid-text insertion (click anywhere to continue dictating at that position) and voice commands (say "delete" to remove highlighted text), VoxBar AI becomes a powerful editing tool — not just a transcription engine.
Technical Specifications
- Model: NVIDIA Canary Qwen 2.5B (NeMo Framework)
- Parameters: ~2.5B (FastConformer encoder + Qwen decoder)
- Architecture: FastConformer encoder + Qwen language model (SALM)
- Languages: English, German, French, Spanish
- Processing: Chunk-based (not streaming)
- Docker: Not required — native Python
- VRAM: ~6-8GB (NVIDIA only)
- License: CC-BY-4.0
VoxBar AI vs. VoxBar Pro
The comparison is straightforward:
- VoxBar Pro — higher accuracy (4.9% WER), true streaming, 13+ languages, but requires Docker on Windows and costs $49.
- VoxBar AI — excellent accuracy, chunk-based (slightly higher latency), 4 languages, but no Docker, simpler setup, and costs $29.
If accuracy is your top priority and you don't mind Docker, choose Pro. If you want easy setup, infinite sessions, and great-but-not-flagship accuracy, choose AI.