cstr / outetts-0.3-1b-GGUF

OuteTTS-0.3-1B GGUF

GGUF quantizations of OuteAI/OuteTTS-0.3-1B for use with CrispASR .

Files

The talker GGUF contains the OLMo-1B LLM. The WavTokenizer decoder GGUF is always needed as a companion file. Q8_0 is recommended for best quality/size balance. Q4_K works but may occasionally shift word boundaries on short utterances.

Note: Q4_K requires CrispASR commit cbe208fa or later, which fixed a WavTokenizer activation bug (SiLU -> GELU in ResNet blocks) that previously degraded quantized output.

Usage with CrispASR

# Build CrispASR
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target crispasr-cli

# Create a speaker profile from reference audio
python tools/reference_backends/outetts_create_speaker.py \
    --audio ref.wav --text "transcript of the reference audio" \
    --out speaker.json

# Synthesize with voice cloning
./build/bin/crispasr --backend outetts \
    -m outetts-0.3-1b-q8_0.gguf \
    --codec-model wavtokenizer-decoder-f16.gguf \
    --voice speaker.json \
    --tts "Hello, how are you today?" \
    --tts-output hello.wav --seed 42

Architecture

• LLM : OLMo-1B (16 layers, 2048 hidden, 16 MHA heads, SwiGLU FFN, parameter-free LayerNorm, RoPE)
• Codec : WavTokenizer single-codebook VQ-GAN (4096 entries, 512-d, 75 tokens/sec)
• Decoder : Vocos backbone (ConvNeXt + pos_net with GroupNorm/GELU ResBlocks) + ISTFTHead
• Output : 24 kHz mono PCM
• License : CC BY 4.0

Conversion

# From HuggingFace model
python models/convert-outetts-to-gguf.py \
    --input OuteAI/OuteTTS-0.3-1B \
    --output outetts-0.3-1b-f16.gguf

python models/convert-wavtokenizer-to-gguf.py \
    --input OuteAI/wavtokenizer-large-75token-interface \
    --output wavtokenizer-decoder-f16.gguf

# Quantize
crispasr-quantize outetts-0.3-1b-f16.gguf outetts-0.3-1b-q8_0.gguf q8_0