fluidaudio_dart 0.3.1

Flutter bindings for FluidAudio: on-device speech-to-text, voice activity detection, speaker diarization and text-to-speech on Apple platforms (CoreML / Apple Neural Engine).

Changelog

0.3.1 — 2026-07-26

Fixes

• Correct the CocoaPods fallback package version and constrain FluidAudio to the supported 0.15.x line.
• Cover explicit microphone and system-audio startup source forwarding.

0.3.0 — 2026-07-22

Capture

• recordToWavPath on FluidMicrophone.start and FluidSystemAudio.start — tees the capture into a WAV file while the same stream feeds every live attachment (streaming ASR, EOU, VAD, frames). Written natively on the capture queue, so audio still never crosses the platform channel; the file records the ASR-grade 16 kHz mono pipeline as 16-bit PCM and is finalized on stop(). Pure sink semantics: recording never starts or stops the capture, survives the system-audio watchdog's chain rebuild, and an unwritable path fails start loudly. File naming, rotation and retention stay with the caller.

0.2.0 — 2026-07-20

Models

• ModelKind.eou — the end-of-utterance streaming model (~450 MB, ms320 chunk variant) joins FluidModels: pre-download with progress via download, probe with isDownloaded, and clear every chunk variant with remove. A FluidEou.create session then starts from cache instead of blocking its caller on a first-run download.

Fixes

• End-of-utterance models now cache under the standard …/FluidAudio/Models/parakeet-eou-streaming/<chunk> layout. FluidAudio 0.15.5's default path resolution doubled the parakeet-eou-streaming segment, so live sessions and model management could never share a cache; they now use one canonical path, and caches written through 0.1.0's doubled layout are migrated in place on first use (no re-download).

0.1.0 — 2026-07-19

Initial release: complete Flutter bindings for FluidAudio 0.15.x (on-device speech on Apple platforms via CoreML / Apple Neural Engine). macOS 14+, iOS 17+, Apple Silicon.

Speech-to-text

• FluidAsr — batch transcription (Parakeet TDT v2/v3), token timings, stateless one-shot semantics.
• FluidStreamingAsr — live sliding-window sessions with volatile/confirmed update streams and strict feed ordering.
• FluidEou — end-of-utterance turn detection with partial/utterance streams.
• FluidCtcVocabulary + configureVocabulary — domain-term boosting (CTC-110M keyword spotter).
• FluidItn — inverse text normalization ("twenty five dollars" → "$25"), graceful no-op when the native library is unavailable.

Audio analysis

• FluidVad / FluidVadStream — Silero voice-activity detection, batch and streaming with probability ticks and speech start/end events.
• FluidDiarizer — offline speaker diarization (VBx) with raw speaker embeddings, speaker database, chunk embeddings, and per-chunk progress.

Text-to-speech

• FluidKokoroTts — Kokoro on the Apple Neural Engine (24 kHz).
• FluidPocketTts — streaming synthesis (80 ms frames) and voice cloning from reference audio.

Capture

• FluidMicrophone — native AVAudioEngine capture, resampled to 16 kHz mono and fanned out natively to attached sessions (audio never crosses the platform channel).
• FluidSystemAudio (macOS 14.4+) — Core Audio process taps capture other applications' audio (all, or targeted PIDs via listAudioProcesses), with a permission preflight for the System Audio Recording prompt.
• Watchdog health streams on both captures (validating → healthy / silent / rebuilding / failed) with a one-shot tap rebuild for late-tappable Electron/Chromium helpers.

Infrastructure

• FluidModels — model download with progress streams, cache inspection, offline mode (models pulled from HuggingFace FluidInference/* on first use).
• FluidAudioConverter — file/sample resampling to 16 kHz mono, WAV encoding.
• Typed FluidAudioException hierarchy carrying real native error messages; Finalizer backstops releasing native models on GC; per-engine plugin lifecycle with teardown on engine detach.

Built as a pigeon platform-channel plugin (shared darwin Swift source, SPM first with a CocoaPods fallback). Verified end-to-end against real CoreML models; four adversarial review rounds (25 confirmed findings fixed).

Known limitation: Qwen3 multilingual ASR is not bound (removed upstream in FluidAudio 0.15.x).