AbstractVoice 0.7

[0.7.0] - 2026-04-06

Added

• ADR 0005: Torch device + dtype selection policy for torch-based TTS/cloning engines.
• Pluggable TTS adapter registry (keeps auto deterministic; enables opt-in heavy engines).
• abstractvoice[audiodit] optional extra and LongCat-AudioDiT integration (TTS + prompt-audio cloning).
• abstractvoice[omnivoice] optional extra and OmniVoice integration (omnilingual TTS + prompt-audio cloning; supports voice design via instruct).
• Shared duration estimation helpers for engines that require explicit duration parameters.
• REPL: /history command to inspect the in-memory LLM message list (what is sent to the provider).
• OmniVoice (REPL/adapter): seed parameter for reproducible voice-design sampling (stable “designed voice” across turns).

Fixed

• AudioDiT TTS now avoids “quiet noise” collapses by retrying with smaller text chunks when outputs are detected as weak.
• AudioDiT TTS now uses a larger default chunk size to reduce voice/pitch drift across multi-sentence utterances.
• AudioDiT TTS duration estimation now follows upstream per-character heuristics (stabilizes voice/pitch across text lengths).
• AudioDiT TTS now defaults to APG guidance (upstream-recommended) for more stable quality.
• AudioDiT TTS now reuses a short “session prompt” (generated prompt-audio + matching prompt-text prefix) to keep a stable speaker identity across multiple /speak calls.
• AudioDiT explicitly rejects /speed changes (upstream doesn’t provide a speed API; attempting speed caused degraded audio).
• Audio playback now falls back to stereo output streams when mono initialization fails (macOS AUHAL robustness).
• Suppressed the noisy PyTorch weight_norm deprecation warning during AudioDiT model load.
• Disabled Transformers progress bars (e.g. “Loading weights”) during AudioDiT model load for cleaner REPL UX.
• Offline-first: Faster-Whisper and AudioDiT now force local-only Hugging Face access when downloads are disabled (avoids HF Hub “unauthenticated requests” warnings).
• Suppressed benign Faster-Whisper mel-extraction RuntimeWarnings (matmul overflow/divide-by-zero) for cleaner REPL output.
• REPL: added /debug and (when enabled) persist each synthesized utterance to untracked/generated_wavs/ and print the path.
• AudioDiT: expand English digits/years into words for more reliable pronunciation (e.g. “5”, “2025”).
• AudioDiT: load model weights in the resolved torch dtype (MPS default fp16; override via ABSTRACTVOICE_TORCH_DTYPE) for better accelerator performance.
• Added engine-agnostic TTS quality presets (fast|balanced|high) via VoiceManager.set_tts_quality_preset(...) (where supported by the active TTS adapter).
• Cloning: set_cloned_tts_quality(...) now persists the preset so it also applies to cloning engines that are loaded later (engines are lazy).
• REPL: selecting an AudioDiT cloned voice now performs a small warm-up to pay the one-time load/compile cost up front (reduces first /speak latency).
• REPL: discard <think>...</think> blocks in LLM responses before printing/history/TTS.
• REPL: LLM chat history is now committed atomically (failed LLM calls no longer leave orphaned user messages that can cause repetition).
• REPL: long pasted prompts no longer crash the clone shortcut path (File name too long); debug WAV filenames are now length-safe.
• Docs: clarify REPL as a demonstrator (minimal LLM client) and recommend AbstractCore for production agent/server integration; refreshed llms.txt / llms-full.txt.
• Docs: clarify AudioDiT language expectations (upstream EN/ZH focus; other languages not guaranteed).
• Docs: expand docs/faq.md with OmniVoice voice design attributes, cross-machine “voice preset” guidance, and fine-tuning/data preparation pointers (plus engine-specific cloning reference recommendations).
• OmniVoice cloning now loads reference audio via soundfile (avoids torchaudio torchcodec dependency for prompt loading).