Changelog #

0.2.2 - 2026-08-06 #

Require mlx ^0.2.0: SmartTurnModel calls setCacheLimit, which first shipped in mlx 0.2.0. The published 0.2.1 declared ^0.1.0 and could resolve an mlx that lacks the symbol.
Changelog correction: the Smart Turn v3.2 entries below shipped in 0.2.1 (they were mis-filed under Unreleased).

Added Smart Turn v3.2 turn detection: SmartTurn.load() resolves the pinned mlx-community/smart-turn-v3 checkpoint, caps MLX's buffer cache at 64 MB, and warms the Metal kernels on 1 s of silence; SmartTurnModel.predictEndpoint scores one ≤ 8 s window of 16 kHz mono audio and returns the sigmoid probability plus a strict probability > threshold verdict.
Added dart run mlx_audio:turn_replay (parity / bench / score) — the replay harness, emitting the same window/probs schema as the ectos Swift harness so the two can be diffed directly or fed to the shared fp32 ONNX reference (tools/smart-turn-parity.py).
Parity: max |Δp| = 0.0000 vs the fp32 ONNX reference and 0.0038 vs the Swift oracle over the committed two-fixture corpus (gate: < 0.02).
Added revision pinning to resolveOrDownloadModel: the new revision parameter (default main, unchanged behaviour) is threaded into both the tree listing and every download URL, so a repo can be resolved at an exact commit.
Added SHA256-verified downloads through the new pinned parameter with PinnedFile / PinnedRepo: each pinned file is hashed as a temp before it is renamed into place, a mismatch raises the new typed HubIntegrityException and leaves nothing behind, and a warm cache is recognised by exact byte count.
Added kSmartTurnWeights — the pinned mlx-community/smart-turn-v3 (Smart Turn v3.2) revision and per-file hash manifest.

Added interface-level STT generation parameters, typed inference progress, and cooperative cancellation.
Added deterministic, idempotent model cleanup through close().
Added true incremental PocketTTS PCM callbacks while preserving the complete TtsOutput return value.

Initial release: on-device speech models for Apple Silicon, built on the mlx package.

Speech-to-text:
- Parakeet-TDT — log-mel front-end → Conformer encoder → RNN-T predict/joint → TDT greedy decode, loading the real mlx-community/parakeet-tdt-0.6b-v2 checkpoint.
- Whisper — mel front-end → encoder → greedy KV-cache decode with >30 s tiled chunking; loads HuggingFace fp32, mlx-whisper fp16, and 4-bit / 8-bit quantized safetensors, with automatic tokenizer fallback by vocab size.
Text-to-speech: PocketTTS — FlowLM backbone + flow-matching latent sampler → Mimi neural codec, synthesizing a 24 kHz waveform from text with eight predefined voices.
DSP / audio: STFT + log-mel front-end, float32 WAV I/O, resampling, and mono-16 kHz loading.
Model management: a Hugging Face hub client that snapshots repos into a local cache (resolveOrDownloadModel, HubProgress).
Eval utilities: transcript normalization and word-error-rate.
CLIs: mlx_audio:transcribe (STT) and mlx_audio:say (TTS).