cstr/Audio8-TTS-Realistic-0.6b-GGUF

Audio8-TTS-Realistic-0.6b — GGUF (ggml-quantised)

GGUF / ggml quantised weights of Edge0/Audio8-TTS-Realistic-0.6b for use with CrispASR.

Audio8-TTS is a Dual-Autoregressive text-to-speech and zero-shot voice cloning model operating at 44.1 kHz. It couples a slow semantic transformer with a fast acoustic transformer and a causal neural vocoder (Snake activation + causal ConvNeXt blocks).

Quantization Details

Standard GGUF quantization corrupts 1D convolutional layers with small kernel taps ($K \le 16$) and discrete acoustic centroids ($k=1024$). CrispASR applies precision-preserving carve-outs:

• Q4_K (1.28 GB): Attention and FFN linear weights quantised to Q4_K / Q6_K; vocoder convs and discrete acoustic embeddings preserved at FP16.
• Q8_0 (1.49 GB): Attention and FFN weights quantised to Q8_0; achieves 1.000000 cosine parity across all stages against PyTorch.
• F16 (1.88 GB): Unquantised IEEE FP16 baseline.

Quickstart with CrispASR

CLI Inference (auto-downloads from registry)

crispasr --backend arktts -m auto --tts "The quick brown fox jumps over the lazy dog." --tts-output speech.wav

Direct GGUF Model Path

crispasr --backend arktts -m arktts-q4_k.gguf --tts "CrispASR delivers high quality text to speech with low latency." --tts-output speech.wav

Performance & Memory

• Autoregressive Generation: 150–220 ms/frame (21.5 frames/s = Real-Time Factor ~0.21x on 8 CPU threads).
• Long Audio Scaling: Chunked causal vocoder decoding bounds peak memory usage strictly under 2.5 GB even for minutes of continuous audio.