Skip to content

v0.15.5

Latest
Latest

Choose a tag to compare

View all tags
@Alex-Wengg Alex-Wengg released this 07 Jul 22:55
v0.15.5
19600a4
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

  • Download stack rebuilt (#765). The monolithic DownloadUtils is replaced by a composable ModelHub with resumable downloads (HTTP Range), byte-level progress, artifact validation before caching, and a shared retry policy that honors Retry-After. Breaking: DownloadUtils → ModelHub (#779).
  • Parakeet unified ASR (#705). Native-Swift mel front-end, word-level timestamps, and lower-latency streaming tiers.
  • Custom vocabulary controls (#647, #702, #724). Per-term CTC thresholds plus opt-in knobs to curb short-term over-firing and disable the acoustic spotter-rescue.

ASR

  • Native-Swift mel front-end fixes iPadOS cold-start zero output (#744)
  • Repair stale/corrupt Nemotron multilingual latin tokenizer on download (#690)
  • Remove chunk-merge seam artifacts in offline transcription (#708), plus word-boundary-safe fallbacks for 3 residual seam-merge drop paths (#759)
  • Fetch parakeet_vocab.json in AsrModels.download (#763)
  • Feature Provider classes for unified Parakeet ASR (#713, @SGD2718)
  • Fix type-checker timeout in NemotronMultilingualFleursBenchmark (#732, @kripskroll)

TTS

  • Shared English text normalization (numbers/times) + initialism helpers (#715); adopted by StyleTTS2 (#720)
  • Auto-chunk long text in KokoroAne high-level synthesize (#717)
  • Fix KokoroAne strided MLMultiArray handling (#730, @XUJiahua)
  • Supertonic: inter-chunk silence 0.3s → 0.05s + --silence flag (#737)

Diarization

  • BNNS-fixed Sortformer v3 models, precision/compute-unit control + cancellation (#728)
  • Deterministic, robust offline VBx re-clustering via K-Means n_init (#735, @testfields)
  • Honor configuration.computeUnits in OfflineDiarizerModels.load (#743, @chadneal)
  • Optional progressHandler on performCompleteDiarization (#753, @thatbin)
  • Honor configuration.computeUnits in OfflineDiarizerModels.load (#743, @chadneal)
  • Optional progressHandler on performCompleteDiarization (#753, @thatbin)
  • Re-embed zero-vote spans instead of arbitrary cluster-0 tie-break (#751, @ComicBit)
  • LS-EEND speaker pre-enrollment bugfixes (#729, @SGD2718)

Download infrastructure (#765)

  • Wave 1: characterization tests + download smoke CI (#770)
  • Wave 2: extract HFClient + RetryPolicy primitives (#771)
  • Wave 3: one paginating HFTreeLister (#772)
  • Wave 4: FileDownloader, ModelCache, ProgressReporter; DownloadUtils becomes a façade (#777)
  • Wave 5: converge fetchHuggingFaceFile onto the shared policy (#778)
  • Wave 6: ModelHub replaces DownloadUtils (#779) — breaking
  • Resumable downloads (#768), byte-level progress (#764), byte-weighted subdirectory progress (#769, @JulianPscheid), artifact validation (#741), preserve cache on cancelled first load (#749, @ComicBit)

VAD

Benchmarks & Docs

  • Fail loudly on missing AMI annotations instead of scoring placeholder ground truth (#754)
  • Fetch AMI corpus from HuggingFace mirror, Edinburgh fallback (#767)
  • Add evoglyph to the Showcase (#755, @lucasmccomb)

New Contributors

@kripskroll, @testfields, @lucasmccomb, @thatbin, @chadneal, @holovchenko

Full Changelog: v0.15.4...v0.15.5

Contributors

kripskroll, JulianPscheid, and 9 other contributors
Assets 2
Loading