add canary stt model (nvidia canary-1b-v2)

[mm65x]

Context

Canary is listed as a planned STT model in the roadmap (#1). NVIDIA's canary-1b-v2 is a top performer on the Open ASR Leaderboard (7.15% avg WER) with support for 25 EU languages plus Russian and Ukrainian, including cross-language translation.

Description

Adds a complete Canary model implementation for mlx-audio's STT pipeline. The model uses a FastConformer encoder (reusing the existing parakeet conformer) paired with a Transformer decoder with cross-attention for autoregressive text generation. Weights are loaded from safetensors converted from NVIDIA's .nemo format.

Changes in the codebase

• mlx_audio/stt/models/canary/canary.py: model class with generate(), sanitize() for NeMo weight mapping, and audio preprocessing
• mlx_audio/stt/models/canary/decoder.py: transformer decoder with self-attention, cross-attention, fixed positional encoding, and KV-cache
• mlx_audio/stt/models/canary/config.py: model configuration dataclasses
• mlx_audio/stt/models/canary/tokenizer.py: sentencepiece tokenizer wrapper with canary prompt format
• mlx_audio/stt/models/canary/__init__.py: module exports
• mlx_audio/stt/utils.py: register "canary" in MODEL_REMAPPING

Changes outside the codebase

None.

Additional information

• Encoder reuses parakeet/conformer.py directly, no duplication
• Tested with canary-1b-v2: English/German transcription and bidirectional translation all produce correct output
• The model has no encoder-decoder projection (Identity) since encoder and decoder dimensions match (1024)
• Decoder uses 8 transformer layers with pre-layer-norm and a final layer norm

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced (e.g., "Closes #...") - addresses Canary item in TTS and STS Models to port to MLX-Audio (Roadmap) #1

[Blaizzy]

Awesome, this was one of the top in our backlog

Could you add a model readme (with inference examples) in the canary folder and link it in the main readme?

[mm65x]

Added a README in the canary folder with usage examples and linked it from the main README's STT table.

[mm65x]

Also, thank you for this library! I'm building a local ASR app for Mac and mlx-audio has been a great option. I've got a couple more models in the pipeline that I'd like to contribute and will open PRs for them too

[lucasnewman]

Please run the formatter: pre-commit run --all so we can clear tests, otherwise looks great!

[mm65x]

done, ran the formatter

[lucasnewman]

🚀