v0.6.0 — Lazy registry, caching, 254 tests

What's New

Lazy Handler Registry (Phase 10)

The single biggest performance improvement: handlers now import only when first used, not at startup.

Before	After
6+ min startup (loads all 21 handlers)	Instant (loads only requested handler)
~2 GB RAM (torch + transformers + outetts)	~200 MB for crispasr-only use
350s test suite	60s test suite

Import Hang Fix (Phase 9)

Root causes identified and fixed:

• config.py: OuteTTS import hangs indefinitely (protobuf compilation) — removed
• utils.py: pygame/sounddevice block on headless machines — deferred to first use
• decoder.py: SNAC model download at import — deferred to first use

New CLI Features

Flag	Description
--backend NAME	Shortcut for CrispASR backends (e.g., --backend kokoro)
--batch	Split input at blank lines, produce numbered output files
--lexicon TSV	Custom word→phoneme mappings for CrispASR
--list-models --check	Probe backend availability
--cache-stats	Show synthesis cache statistics
--cache-clear	Clear cached synthesis results
--rate-limit N	Max requests/min/IP for server (default 10)

Server Improvements

• Threaded: Concurrent requests via ThreadingMixIn
• Rate limiting: Token bucket per IP (429 on exceed)
• Synthesis caching: Hash-based LRU, serves identical requests from cache
• Enhanced /health: Reports loaded handlers, memory RSS, registered backends
• Content-Disposition: Proper download filename headers

Performance Optimizations (Phase 8)

• Reduced file I/O in watermark pipeline (4 ops → 3)
• Lazy watermark model loading (~200MB deferred until first synthesis)
• Streaming synthesis concurrency limit (Semaphore(4))
• Subprocess stdout → DEVNULL (only stderr captured)

Test Coverage (Phase 11)

254 tests passing in ~60s:

• 30 new tests for cache, crossfade, lazy registry, rate limiting, config validation, --backend shortcut
• New test files: test_cache.py, test_server.py

Audio Utilities

• crossfade_segments(): Linear crossfade for concatenating chunked audio
• Config validation at startup with clear warnings
• Batch error recovery (failed paragraphs skipped, summary at end)

v0.5.0 — 14 CrispASR backends, TADA inline cloning

What's New

4 New CrispASR TTS Backends

CrispTTS now exposes 14 native C++ TTS backends via CrispASR (up from 10):

Backend	Type	Sample Rate	Voice Cloning
BananaMind-TTS	Tacotron-lite + HiFi-GAN	22050 Hz	No
Dots.TTS	Qwen2.5 LLM + DiT + BigVGAN	48000 Hz	Yes (CAM++)
CosyVoice3	Multi-GGUF (LLM+flow+CAM+++HiFT)	24000 Hz	Yes
CSM/Sesame	Llama-3.2 + Mimi codec	24000 Hz	Yes

TADA Inline Voice Cloning

New --ref-text flag enables one-command voice cloning with TADA — no separate --make-ref step needed:

python main.py --model-id crispasr_dots_tts \
  --input-text "New speech" --german-voice-id ref.wav --ref-text "transcript of ref audio" \
  --i-have-rights --output-file cloned.wav

Expanded Model Parameters

12 new --model-params keys for per-request tuning: top_k, min_p, cfg_scale, do_sample, num_candidates, num_steps, noise_temp, noise_scale, noise_w, speaker_id, max_speech_tokens.

Performance Optimizations

• Lazy watermark loading: Neural backends (WavMark/AudioSeal ~200MB) now load on first synthesis, not at CLI startup. --list-models and --help are instant.
• Reduced file I/O: Post-synthesis watermark + metadata pipeline uses fewer read/write cycles.
• Streaming concurrency limit: Semaphore(4) prevents unbounded thread growth.
• Subprocess optimization: Stdout no longer buffered (only stderr captured for errors).
• Server fixes: File handle leak fixed, Content-Disposition header added.

Other Changes

• --no-spoken-disclaimer flag to skip AI disclosure prefix on cloned audio
• Voice-cloning detection expanded for dots-tts, cosyvoice3, csm, tada backends
• Release workflow fixed to handle pre-existing releases
• 224 tests passing (py3.10/3.11/3.12 + ruff)

v0.4.0 — EU AI Act Art. 50 compliance

What's New

EU AI Act Article 50 Compliance (deadline: Aug 2, 2026)

Watermark embedding on all outputs — the critical gap where watermark functions existed but were never called is now closed. Every synthesized audio file gets:

• Audio watermark: WavMark (MIT, preferred), AudioSeal, or spread-spectrum (always-on fallback)
• File metadata: WAV LIST/INFO, MP3 ID3v2, FLAC Vorbis comments, Opus OGG tags — all declare AI_GENERATED=true
• C2PA content credentials: cryptographically signed provenance manifests (when cert/key configured)

Voice Cloning Safety

• Server API consent gate: POST /v1/audio/speech now requires "i_have_rights": true for cloning models; returns 403 with explanation otherwise
• Expanded detection: CrispASR cloning backends (vibevoice, indextts, voxcpm2, qwen3-tts, f5-tts) now caught by keyword matching
• Persistent audit log: consent attestations written to ~/.cache/crisptts/consent_audit.log (not just stderr)
• Spoken disclaimer: uses CrispASR kokoro (local) first, Edge TTS fallback, beep marker last resort

New CrispASR TTS Backends

• crispasr_f5_tts — F5-TTS flow-matching, voice cloning, 24 kHz (Apache 2.0)
• crispasr_melotts — MeloTTS VITS2, 44.1 kHz (MIT)
• crispasr_piper — Piper VITS via C++, 250+ community voices, 30+ langs

WavMark Neural Watermark (MIT license)

Added as the preferred neural watermark backend — fully MIT licensed (code + model weights), unlike AudioSeal (CC-BY-NC weights). Priority chain: WavMark → AudioSeal → CrispASR GGUF → spread-spectrum.

Tests

212 tests passing across Python 3.10–3.12. 13 new tests covering watermark embedding, voice cloning detection, audit logging, FLAC/Opus metadata, and multi-sample-rate dispatch.

Install optional features

pip install crisptts[watermark-mit]   # WavMark (MIT neural watermark)
pip install crisptts[metadata]        # FLAC/Opus metadata via mutagen
pip install crisptts[provenance]      # C2PA content credentials

v0.3.0

What's Changed

• fix(ci): unblock Dependabot dependency-graph submission by @CrispStrobe in #1
• chore(release): prep v0.3.0 — dependency triage + release workflow by @CrispStrobe in #9

New Contributors

• @CrispStrobe made their first contribution in #1

Full Changelog: https://github.com/CrispStrobe/CrispTTS/commits/v0.3.0