github FluidInference/FluidAudio v0.18.1
v0.18.1 — LFM3-Audio, Phonon-2.5, vocabulary thresholds

latest release: v0.18.2
2 hours ago

New models

  • LFM3-Audio-2B (#1076, #1077, #1078): ASR, live ASR and TTS on Core ML and Core AI. Cold loads are hardened for low disk.
  • Phonon-2.5 multilingual ASR on the GPU (#1075).

ASR

  • Explicit streaming token revisions with a monotonic stable prefix (#1035).
  • Cancelled streaming recognition waits for its task to exit (#1033).
  • The CTC mel spectrogram runs on the CPU, fixing all-blank CTC on M1/M4 with macOS 27.0.1 (#1079).

Vocabulary boosting: session thresholds (VocabularyBoostingSession.Thresholds: cbw, minSimilarity, marginSeconds) and a detectedTerms floor (#1080), and names, possessives and neighbouring words are preserved (#1041).

TTS: Pocket TTS cuts oversized sentences at more kinds of punctuation and balances word splits (#935).

What's Changed

  • feat(asr,tts): LFM3-Audio-2B: ASR, LiveASR and TTS on Core ML and Core AI by @Alex-Wengg in #1076
  • perf(tts): trim LFM3-Audio TTS host work; report TTS process CPU by @Alex-Wengg in #1077
  • feat(asr): add Phonon-2.5 multilingual GPU support by @pHequals7 in #1075
  • fix(lfm3-audio): harden cold loads on a low disk by @Alex-Wengg in #1078
  • fix(ctc): run the CTC MelSpectrogram on the CPU by @Alex-Wengg in #1079
  • feat(vocab): session thresholds and a detectedTerms floor (#967) by @Alex-Wengg in #1080
  • fix(asr): wait for cancelled streaming recognition to exit by @achembarpu in #1033
  • fix(asr): Preserve names, possessives and surrounding words by @achembarpu in #1041
  • feat(asr/streaming): expose explicit token revisions with a monotonic stable prefix by @johnny4young in #1035
  • feat(tts/pocket): cut oversized sentences at more kinds of punctuation and balance word splits by @arvindvenkataramani in #935

New Contributors

Full Changelog: v0.18.0...v0.18.1

Don't miss a new FluidAudio release

NewReleases is sending notifications on new releases.