aoiandroid/parakeet-tdt-ja-coreml

Parakeet TDT-CTC 0.6B Japanese — Core ML (Streaming ASR)

Core ML conversion of nvidia/parakeet-tdt_ctc-0.6b-ja for SlidingWindow streaming ASR on iOS/macOS (TranslateBlue / FluidAudio).

Streaming ASR contract

This repo is not a batch-only file transcription bundle. It targets live microphone ASR:

Aspect	Specification
Architecture	SlidingWindow pseudo-streaming (stateless 15s encoder + overlapping windows)
Runtime config	SlidingWindowAsrConfig.streaming — 11s chunk, 2s left/right context
Decoder	Decoderv2 with U=1 LSTM state I/O (state carried across windows)
Jointer	Jointerv2 — TDT duration bins (max 4 frames for ja)
Encoder window	Fixed 15s mel input ([1, 80, 1501]). Short windows (e.g. 5s) are unsupported.
Vocab	3072 BPE tokens, blank id 3072
CTC	CtcDecoder.mlpackage is tier-2 failover only (raw logits, no log_softmax)

Not included: cache-aware true streaming (Parakeet EOU / Nemotron), fused FullPipeline / MelEncoder batch paths.

Artifacts (aoiandroid lowercase layout)

File	FluidAudio name	Role
preprocessor.mlpackage	Preprocessor	16 kHz mono → mel
encoder.mlpackage	Encoder	mel → encoder_output
decoder.mlpackage	Decoderv2	LSTM decoder with state
joint.mlpackage	Jointerv2	TDT joint step
vocab_ja.json	vocab.json	SentencePiece vocabulary
CtcDecoder.mlpackage	CtcDecoder	CTC tier-2 failover

Conversion

Pipeline: FluidInference/mobius
Script: models/stt/parakeet-ctc-0.6b-ja/coreml/conversion/export-tdt-ja-streaming.py

cd mobius/models/stt/parakeet-ctc-0.6b-ja/coreml
uv sync
uv run python conversion/export-tdt-ja-streaming.py --output-dir ./build
uv run python conversion/export-full-pipeline.py --output-dir ./build --no-fused

Validation

• Streaming: fluidaudiocli transcribe <wav> --streaming --model-version tdt-ja --model-dir ./build
• Accuracy: JSUT TDT CER via fluidaudiocli ja-benchmark --dataset jsut --samples 500

License

Converted weights follow the upstream NVIDIA model license. Model card metadata: CC-BY-4.0.