miocodec-v2-44k-GGUF

MioCodec-25Hz-44.1kHz-v2 GGUF

GGUF conversions of Aratako/MioCodec-25Hz-44.1kHz-v2 (MIT license, 133M params) for use with CrispASR.

MioCodec is a lightweight neural audio codec: encode 44.1 kHz audio to 25 Hz discrete tokens (12800 vocab, 341 bps), decode tokens + speaker embedding back to waveform.

Files

All GGUFs contain the decoder-only weights (FSQ quantizer + wave_prenet + wave_decoder + wave_post_net + SnakeBeta upsampler + ISTFTHead). The WavLM encoder is separate.

Architecture

tokens(T) -> FSQ(T, 768) -> wave_prenet 6L Transformer(T, 512)
  -> ConvTranspose1d 2x -> interpolate -> ResNet x2
  -> wave_decoder 8L Transformer+AdaLN-Zero(T', 512) -> ResNet x2
  -> SnakeBeta Upsampler 9x -> ISTFTHead(394) -> 44.1kHz waveform

• FSQ: Finite Scalar Quantization, levels=[8,8,8,5,5], 12800 codebook
• Transformer: RoPE (normal/adjacent-pair), SwiGLU FFN, windowed attention
• AdaLN-Zero: Speaker conditioning via 128-dim global embedding
• SnakeBeta: x + exp(-beta) * sin^2(exp(alpha) * x) periodic activation
• ISTFT: exp(log_mag) + cos/sin(phase) -> iFFT overlap-add, "same" padding

Parity (F16 vs Python F32 reference)

Usage with CrispASR

# Diff harness (verify per-stage parity)
crispasr-diff miocodec miocodec-v2-44k-f16.gguf ref.gguf audio.wav

Conversion

python models/convert-miocodec-to-gguf.py \
  --model Aratako/MioCodec-25Hz-44.1kHz-v2 \
  --output miocodec-v2-44k-f16.gguf --dtype f16

# Quantize
crispasr-quantize miocodec-v2-44k-f16.gguf miocodec-v2-44k-q8_0.gguf q8_0
crispasr-quantize miocodec-v2-44k-f16.gguf miocodec-v2-44k-q4_k.gguf q4_k

License

MIT (same as the original model).

Provenance and EU AI Act Art. 53 note

• Upstream model: Aratako/MioCodec-25Hz-44.1kHz-v2 — published by Aratako.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.