BricksDisplay/codec.cpp-gguf

codec.cpp GGUFs

Neural audio codecs for codec.cpp: turn speech into discrete tokens and back, offline on CPU or GPU.

Codec	Sample rate	Token rate	Files
IBM DAC.speech v1.0	24 kHz	75 frames/s	ibm-research--DAC.speech*.gguf
NVIDIA NeMo Nano Codec	22.05 kHz	12.5 frames/s (0.6 kbps)	nemo-nano-codec-22khz-0.6kbps-12.5fps*.gguf
WavTokenizer large (speech)	24 kHz	75 tokens/s	wavtokenizer-large-speech-75tokens*.gguf

Each codec comes in several precisions: the file without a suffix is F16 (recommended), -F32 is full precision, -Q8_0 / -Q5_K_M / -Q4_K_M are smaller and faster at some quality cost.

Run

hf download BricksDisplay/codec.cpp-gguf ibm-research--DAC.speech.gguf --local-dir codecs

./build/codec-cli encode --model codecs/ibm-research--DAC.speech.gguf --in speech.wav --out codes.npy
./build/codec-cli decode --model codecs/ibm-research--DAC.speech.gguf --codes codes.npy --out decoded.wav

Python (pip install . in a codec.cpp checkout; see python/README.md):

import codec_cpp, soundfile as sf

ctx = codec_cpp.Model("codecs/ibm-research--DAC.speech.gguf").context()
pcm, sr = sf.read("speech.wav", dtype="float32")
codes = ctx.encode(pcm, sr)          # int32 (n_frames, n_q)
audio, sr = ctx.decode(codes)

Build instructions and GPU flags are in the codec.cpp README.

These codecs are also the audio side of TTS models: there a llama.cpp LM backbone writes the codes and codec.cpp turns them into audio (see the *-GGUF TTS repos of this organization).

License

Each codec keeps its original license: DAC.speech CDLA-Permissive-2.0, NeMo Nano Codec the NVIDIA Open Model License, WavTokenizer MIT.