v0.8.49

v0.8.49 — Audio8-TTS (ArkTTS), LID Standalone/Aggregation, and Qwen3 Context

Speech synthesis

• Full backend port and integration for Audio8-TTS (ArkTTS 0.6B) (Edge0/Audio8-TTS-Realistic-0.6b):
  • Dual-autoregressive text-to-speech architecture coupling a 28-layer slow semantic transformer with an 8-layer fast acoustic subnetwork generating 10 codebooks per frame at 44.1 kHz.
  • Causal neural vocoder with Snake activations and ConvNeXt residual blocks.
  • Precision-preserving GGUF quantization: attention and FFN linear weights quantised to Q4_K / Q8_0 while keeping 1D conv kernels ($K \le 16$) and discrete acoustic centroids ($k=1024$) at FP16 to guarantee bit-exact arithmetic without acoustic hash noise.
  • Verified 1.000000 cosine parity across all 7 stages on crispasr-diff against the PyTorch reference implementation.
  • Performance optimizations:
    • FASTCONV $K=1$ bypass eliminating im2col transposition and contiguous memory buffer copies across 12 audio-rate layers.
    • Dual-transformer in-place KV caching.
    • Chunked causal vocoder decoding with causal history overlap, bounding peak RAM to <2.5 GB for arbitrary audio lengths (3+ minutes) where monolithic graphs require >160 GB.
  • Official GGUF weights published to Hugging Face repository cstr/Audio8-TTS-Realistic-0.6b-GGUF and registered in crispasr_model_registry.
  • Verified 0.0% Word Error Rate (WER) across TTS-to-ASR roundtrips.

Recognition and CLI

• Support free-form transcription context in Qwen3-ASR (#559, thanks @pszemraj).
• Language Identification (LID) improvements (#560):
  • Standalone --detect-language (-dl) action performs language detection and exits cleanly without requiring an ASR model to be loaded or executed.
  • Dominant language detection across full recordings.
  • Automatic provider fallback to Whisper when dedicated LID models are not specified.

Streaming and context

• Adaptive strided audio context gating for Hikari and Parakeet streaming inference.
• Attention mask caching in Moonshine to eliminate redundant mask construction during streaming chunk processing.
• Synchronize audio_ctx and hikari_refresh_stride across Go, Dart, Rust, Python, JavaScript, and C ABI.
• Formatted Dart session setter lookupFunction to pass strict CI format checks.

Translation and tools

• Madlad translation: sharded safetensors conversion, low-RAM mmap scratch buffer allocation, and Kaggle quantizations for Madlad-7B and Madlad-10B.
• Cleaned and migrated historical maintenance handoffs and benchmark receipts into docs/historical/.