cstr/miotts-0.6b-GGUF

MioTTS-0.6B GGUF (v2 codec — 44.1kHz)

GGUF conversion of Aratako/MioTTS-0.6B + Aratako/MioCodec-25Hz-44.1kHz-v2 for CrispASR.

Single GGUF containing the LLM (Qwen3, 28 layers, 1024d) and the MioCodec-v2 waveform decoder with UpSampler. Produces 44.1 kHz mono audio.

Files

File	Size	Description
miotts-0.6b-f16.gguf	1.4 GB	Full precision (F16 weights)
miotts-0.6b-q8_0.gguf	793 MB	LLM Q8_0 + codec F16
miotts-0.6b-q4_k.gguf	502 MB	Recommended — LLM Q4_K + codec F16
tokenizer.json	~14 MB	Qwen3 BPE tokenizer (place next to GGUF)
en_female.emb.gguf	<1 KB	English female voice preset
en_male.emb.gguf	<1 KB	English male voice preset
jp_female.emb.gguf	<1 KB	Japanese female voice preset
jp_male.emb.gguf	<1 KB	Japanese male voice preset

Usage

crispasr --backend miotts -m miotts-0.6b-q4_k.gguf \
  --voice en_female.emb.gguf \
  --tts "Hello world, how are you today?" \
  -of output

Important: A voice preset (--voice or -emb) is required for good audio quality. Without it, the codec produces unintelligible audio.

The tokenizer.json must be in the same directory as the GGUF file.

Verified ASR Roundtrip

Quantization	Size	Input	ASR Output
Q4_K+F16	502 MB	"The quick brown fox jumps over the lazy dog."	"The quick brown fox jumps over the lazy dog." ✅
F32	2.6 GB	"Hello world, how are you today?"	"Hello world, how are you today? What?" ✅
F16 (uniform)	1.4 GB	—	Codec too noisy (SnakeBeta precision) ❌

The mixed quantization (LLM=Q4_K, codec=F16) is critical — uniform quantization degrades the UpSampler's SnakeBeta activation.

Architecture

• LLM: Qwen3ForCausalLM (28L, 1024d, GQA 16/8, head_dim=128, vocab=164480)
• Codec: MioCodec-25Hz-44.1kHz-v2 (FSQ → wave_prenet → conv_upsample → ResNet → AdaLN-Zero decoder → ResNet → UpSampler with SnakeBeta → iSTFT)
• Output: 44.1 kHz mono PCM

License

Apache 2.0 (MioTTS-0.6B is Qwen3-based). MioCodec weights under MIT.

Credits

• Model: Aratako
• GGUF conversion + C++ port: CrispASR
• Voice presets: mmnga/mio-tts-cpp