[RahulRachuri]

Closes #163

Type of Change

• Bug fix
• Small feature (agreed in an issue)
• New model (export recipe only)
• New model (export recipe and new runtime support)
• Documentation update

Motivation and Context

Adds NVIDIA's English parakeet-tdt-0.6b-v2 (CC-BY-4.0) alongside v3, as discussed in #163. No Swift changes: the runtime reads the vocabulary, blank ID and durations from the bundle's metadata.

NVIDIA publishes v2 only as a .nemo file, which transformers cannot load directly, so the recipe uses a conversion to the Hugging Face format at rahulrachuri/parakeet-tdt-0.6b-v2. If you would rather not list a personal repository, I am happy to change the source. Exported assets credit NVIDIA's original model.

export.py previously hardcoded v3's name, author and source URL in the asset metadata. These now come from a per-model table that also supplies the --model choices. v3's metadata is unchanged.

Sample runtime output

swift run -c release speech-recognizer --model exports/parakeet-tdt-0.6b-v2_float16_static --audio-path clip180.wav

This is a LibriVox recording. All Librivox recordings are in the public domain. For more information or to volunteer, please visit Librivox.org. This reading by Stuart Wills. Moby Dick By Herman Melville Chapter 3 The Spouter Inn Entering that gable-ended spouter inn, you found yourself in a wide, low, straggling entry, with old-fashioned wainscots, reminding one of the bulwarks of some condemned old craft.

Performance

speech-recognizer on an M4 Pro, macOS 27.2, with a static bundle exported with --audio-seconds 180, transcribing the first 180 s of LibriVox's Moby Dick chapter 3. Process time is the last two of three runs, including model load.

dtype	Bundle size	Decode	Process time
float16	1.2 GB	1.2 s	2.3 s
float32	2.3 GB	1.6 s	3.2 s

Testing

• Tested locally
• Added/updated tests
• Added/updated docs

🤖 Generated with Claude Code