playstonex/Parakeet-TDT-v3-CoreML-INT8

Parakeet TDT v3 — CoreML INT8

CoreML conversion of NVIDIA Parakeet-TDT 0.6B v2 with INT8-quantized encoder for Apple Neural Engine acceleration.

Models

Model	Description	Compute	Quantization
encoder.mlmodelc	FastConformer encoder (24L, 1024 hidden)	CPU + Neural Engine	INT8 palettized
decoder.mlmodelc	LSTM prediction network (2L, 640 hidden)	CPU + Neural Engine	FP16
joint.mlmodelc	TDT dual-head joint (token + duration logits)	CPU + Neural Engine	FP16

Additional Files

File	Description
vocab.json	SentencePiece vocabulary (1024 tokens)
config.json	Model configuration

Notes

• INT8 vs INT4: INT8 uses 8-bit palettization for the encoder, offering higher accuracy than INT4 at the cost of ~2x encoder weight size.
• Mel preprocessing is done in Swift using Accelerate/vDSP (not CoreML) because torch.stft tracing bakes audio length as a constant, breaking per-feature normalization for variable-length inputs.
• Encoder uses EnumeratedShapes (100–3000 mel frames, covering 1–30s audio) to avoid BNNS crashes with dynamic shapes.

Usage

Used by speech-swift ParakeetASR module:

let model = try await ParakeetASRModel.fromPretrained(modelId: ParakeetASRModel.int8ModelId)
let text = try model.transcribeAudio(samples, sampleRate: 16000)

---

---

• Guide: soniqo.audio/guides/parakeet
• Docs: soniqo.audio
• GitHub: soniqo/speech-swift