v0.9.5

v0.9.5 — iOS UIBackgroundModes fix (App Store upload)

v0.9.4

v0.9.4 — iOS altool --apple-id fix (App Store upload)

v0.9.3

v0.9.3 — iOS export IPA-name glob fix (App Store upload)

v0.9.2

v0.9.2 — iOS App Store signing fix (manual profiles) + glint_audio re…

v0.9.1

v0.9.1 build 71 — iOS icon fix

v0.9.0

v0.9.0 — CrispASR 0.8.7 + CrispEmbed 0.13.0 integration + store readi…

main

feat: iOS App Store signing pipeline (API-key-driven)

Following the proven appstore.md playbook (CrispChess/CrispSudoku/
CrispDeck pattern):

- ExportOptions.plist: method app-store → app-store-connect
- pbxproj: delete 3x stale CODE_SIGN_IDENTITY[sdk=iphoneos*]
  overrides that conflict with Automatic signing (Step 4 gotcha)
- .gitignore: add .appstoreconnect/ and *.p8
- release.yml: conditional signed path when APPSTORE_API_KEY_P8
  secret is set — archive → export → validate → upload to App
  Store Connect. Falls back to unsigned IPA when secret absent.

To activate: set these repo secrets in GitHub:
  APPSTORE_API_KEY_P8  (base64 of AuthKey_9RMU3C7422.p8)
  APPSTORE_APP_ID      (numeric, from App Store Connect)

Human steps still required:
  1. Create app record in App Store Connect (browser)
  2. App Privacy nutrition label (browser)
  3. Set APPSTORE_APP_ID secret after step 1

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

v0.8.8

feat: CrispASR 0.8.7 + CrispEmbed 0.13.0 integration sweep

New capabilities:
- Reranker integration in semantic search (cross-encoder re-scoring)
- BidirLM-Omni cross-modal text↔audio search
- Chatterbox emotion tag buttons ([laugh]/[whispering]/[angry])
- TADA standalone re-align timestamps
- OCR service (pix2tex, HMER, BTTR, PosFormer, Granite Vision, DeepSeek-OCR2)
- Scan preprocessing pipeline (deskew, crop, whitening, denoise)
- Wyoming protocol server for Home Assistant STT
- Streaming segment callbacks for LLM-ASR partial results
- LoRA hot-swap Dart FFI binding
- WASM CrispEmbed IndexedDB model caching
- imatrix-quantized embed model defaults

Catalog: 369 entries from 120 HF repos (6 OCR, 3 reranker, 4 embed added)
Tests: 1104 pass, 23 skip, 0 fail

v0.8.7

chore: bump version to 0.8.7+68 for release

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

v0.8.6 — CrispASR 0.8.x parity: 7 new backends + TTS/ASR controls

v0.8.6 — CrispASR 0.8.x parity: 7 new backends + TTS/ASR controls

Catches up with 246 CrispASR commits since v0.8.0: new model catalog entries, new TTS sampling controls, ASR beam/hotword tuning, and server/CLI additions.

New model catalog entries

Model	Type	Size	Description
DoTs-TTS	TTS	~4.8 GB (3 GGUFs)	rednote-hilab 2B continuous AR + flow-matching + BigVGAN, 48 kHz, CAM++ voice cloning
Higgs-STT	ASR	~2.3 GB	bosonai Whisper-v3 encoder + Qwen3-1.7B decoder, internal chunking, beam search, --ask
ARK-ASR 3B	ASR	~2.2 GB	Whisper-RoPE encoder + Qwen2.5-3B decoder, 19 languages
MOSS-Transcribe 2B	ASR	~1.6 GB	Qwen3-Omni encoder + Qwen3-1.7B decoder, native punctuation, streaming
Gemma4-E4B	ASR	~4.1 GB	Larger Gemma4 decoder (42L×2560), 140+ languages
ReazonSpeech NeMo v2	ASR	~704 MB	Japanese RNNT (619M params), Q8_0
Parakeet-CTC 1.1B JA	ASR	~1.2 GB	Japanese FastConformer-CTC, Q8_0

All models auto-download from HuggingFace and appear in the model picker.

New TTS controls (Synthesize screen → Advanced)

• Top-K sampling width slider (0–100)
• Stochastic sampling toggle (greedy vs. stochastic)
• Acoustic candidates slider for TADA/chatterbox/kokoro ranking (0–10)
• Noise temperature slider for kokoro/vibevoice (0.0–2.0)
• G2P dictionary path for kokoro/vibevoice/speecht5
• All controls localized in English, German, and Chinese

New ASR controls (Transcription → Advanced Options)

• Beam width slider (0–20, shown when beam search is enabled) — controls setBeamSize() for whisper/canary/cohere/higgs-stt/ark-asr
• Hotwords boost slider (0.0–5.0, shown when hotwords are entered) — CTC/TDT keyword biasing factor for granite/parakeet

Server + CLI

• diarized_json response format — POST /v1/audio/transcriptions with response_format=diarized_json groups segments by speaker
• diarize-speakers CLI alias for the diarize command

CrispASR 0.8.x fixes (automatic on dylib rebuild)

• Granite long-audio: chunk-context spaceless text + missing passages fix
• Parakeet overlapping-window merge for long audio
• VibVoice GPU crash on long inputs
• TADA codec 805 MB dead masks dropped + Vulkan fixes
• Qwen3-TTS OOM fix on long reference audio
• LFM2-Audio GPU embed+decode CUDA fix
• Vulkan f16 GQA-REPEAT guard

Tests

47 new unit tests (1017 total, 0 failures).