v0.18.0 — Nemotron Unified ASR, Nemotron 3.5 Transcribe Lite, Bark 3.0, Canary translation

New ASR models

• Nemotron Unified ASR 0.6B (#1071): NVIDIA's multilingual Parakeet Unified, 25 European languages, offline and streaming (0.56 s / 1.12 s / 2.08 s latency), on the existing Unified managers via UnifiedConfig(family: .nemotronMultilingual(.de)). CLI: transcribe --model-version nemotron-unified --unified-language de. Models: FluidInference/nemotron-labs-unified-asr-0.6b-coreml.
• Nemotron 3.5 Transcribe Lite (#1069, #1070): the ASR encoder + CTC timestamp head of NVIDIA's Nemotron 3.5 Transcribe (no LLM), words with 40 ms start times plus Nemotron-3-Diarization speakers running in parallel on the Neural Engine. A 49.5-min meeting in ~5.5 s on an M5 Pro. CLI: nemotron35-transcribe.
• Underdog Bark 3.0 + Bark 3.0 Clean (#1054): AsrModelVersion.bark3 / .bark3Clean.
• Canary speech translation (#833, #1030): en ↔ 24 European languages via prompt, with pause-aligned long-form decoding.
• Phonon-2 encoder ladder and pause-cut windows: 480× → 650× on long audio (#1019).

TTS: Pocket TTS small models (18m/7m/2m) with the FlowLM on the Neural Engine (#1051), Gradium Phonon 31m (#1053), Paradee-8M v1.2 with four voices and British G2P (#1073), Supertonic 3 v2 bundles (#1002), Kokoro ANE fixes (#1028, #1038, #1048).

Diarization: per-chunk and per-speaker embeddings, embedSpans, configurable segmentation batch size (#992, #1014, #1015, #1018).

Package: TTS and Diarizer are opt-out traits (#1001); the library builds for tvOS 17+ (#1026).

What's Changed

• feat(tts/supertonic3): v2 CoreML bundles fix padding leak at chunk ends by @Alex-Wengg in #1002
• fix(asr/nemotron-multilingual): publish only rescue-settled text in the partial callback by @JulianPscheid in #994
• feat(diarizer): expose per-chunk embeddings from DiarizerManager by @ertra in #992
• fix(asr/nemotron-multilingual): skip unchanged partial callbacks by @JulianPscheid in #1005
• fix(wordspotting): zero CTC keyword-spotter audio padding by @mvanhorn in #1003
• Avoid unnecessary CTC vocabulary inference by @achembarpu in #993
• fix(asr): take seam-token timing from the window that heard it mid-window by @maboa in #975
• docs(tts): document Kokoro English dictionary versus BART comparison by @Alex-Wengg in #1006
• docs(tts): document Kokoro ANE A12 memory and routing limitations by @mvanhorn in #1009
• docs(tts/supertonic3): weights are BigScience Open RAIL-M, not Apache-2.0 by @Alex-Wengg in #1016
• docs(tts/luxtts): document provenance and licenses of the G2P resources by @Alex-Wengg in #1017
• feat(diarizer): configurable offline segmentation batch size by @Toasty360 in #1014
• feat(asr): Phonon-2 encoder ladder + pause-cut windows — 480× → 650× on long audio by @Alex-Wengg in #1019
• fix(asr/cohere): 256-slot v3 decoder + runtime cache length; cut seams after the matched run by @Alex-Wengg in #1024
• fix(tts): honour the caller's directory for G2P and lexicon assets by @Alex-Wengg in #1025
• docs: add inlaut to showcase by @tobymarks in #1008
• feat(package): TTS and Diarizer opt-out traits (#990) by @Alex-Wengg in #1001
• fix(asr): window progress count matches the windows actually decoded by @Alex-Wengg in #1020
• feat: build the FluidAudio library for tvOS 17+ by @Alex-Wengg in #1026
• feat(diarizer): clean per-speaker embeddings for enrollment by @akre54 in #1018
• fix(tts/kokoro-ane/zh): make g2pW polyphone disambiguation run against the shipped bundle by @Alex-Wengg in #1028
• feat(asr/canary): speech translation (en ↔ 24 European languages) via prompt by @Alex-Wengg in #833
• ci: fix the two main-breaking Build and Test jobs by @Alex-Wengg in #1029
• feat(asr/canary): pause-aligned long-form decoding via VAD for speech translation by @Alex-Wengg in #1030
• docs: add Shiyin to showcase by @juinwei7 in #1036
• fix(tts/kokoro-ane-v3): allocate inputs with tail slack to stop NaN audio on macOS 27 by @Alex-Wengg in #1038
• Reduce CTC keyword-scoring memory and CPU work by @achembarpu in #1034
• docs: add Dictaty to showcase by @comlui in #1042
• fix(asr/cohere): default encoder and decoder to ANE, per-model compute units by @Alex-Wengg in #1046
• feat(diarizer): embed known spans with OfflineDiarizerManager.embedSpans by @Toasty360 in #1015
• fix(download): stop kokoro g2p-only from sweeping every root-level .json by @Alex-Wengg in #1047
• fix(tts/kokoro-ane-v3): skip source .mlpackage checks once the compiled bundle exists by @Alex-Wengg in #1048
• fix(asr/tdtJa): turn off the word-boundary splice guard for the unspaced Japanese vocabulary by @ertra in #1032
• feat(tts/pocket): small models (18m/7m/2m) on the Neural Engine via one MLState package by @Alex-Wengg in #1051
• feat(tts/pocket): Gradium Phonon 31m on the small-model host + Phonon text normalizer by @Alex-Wengg in #1053
• docs: add Helicon to showcase by @serhii-kucherenko in #1060
• fix(examples/pocket-small-demo): session race, stale highlight, dead session, pending leaks by @Alex-Wengg in #1052
• docs: add Tamka to showcase by @piedol in #1066
• Add Barkenson to Showcase by @ziz11 in #1050
• fix(tts/pocket): emit for uncovered scalars when the model has no byte fallback by @Alex-Wengg in #1068
• Surface CTC vocabulary rescoring decisions on the public ASR results by @kevin-nous in #929
• feat(asr): Nemotron 3.5 Transcribe Lite (encoder + CTC words) with Nemotron-3-Diarization speakers by @Alex-Wengg in #1069
• fix(asr/nemotron35-lite): check the OS before downloading; reject wrong-size mel tables by @Alex-Wengg in #1070
• feat(asr/unified): Nemotron Unified ASR 0.6B (25 European languages) on the Unified managers by @Alex-Wengg in #1071
• feat(asr): Underdog Bark 3.0 + Bark 3.0 Clean (AsrModelVersion.bark3 / .bark3Clean) by @Alex-Wengg in #1054
• feat(tts/paradee): Paradee-8M v1.2 with four voices and British G2P by @Alex-Wengg in #1073

New Contributors

• @mvanhorn made their first contribution in #1003
• @achembarpu made their first contribution in #993
• @Toasty360 made their first contribution in #1014
• @tobymarks made their first contribution in #1008
• @akre54 made their first contribution in #1018
• @juinwei7 made their first contribution in #1036
• @comlui made their first contribution in #1042
• @serhii-kucherenko made their first contribution in #1060
• @piedol made their first contribution in #1066
• @ziz11 made their first contribution in #1050
• @kevin-nous made their first contribution in #929

Full Changelog: v0.17.7...v0.18.0