BlueCodec
The 44.1 kHz speech autoencoder of BlueTTS BlueV3: audio to 24-dimensional continuous latents at ~86 Hz and back. It is the codec only, not a text-to-speech model.
One file: model.safetensors holds the encoder (encoder.*) and the decoder (decoder.*), EMA weights at step
380k, with the architecture config stored in the file's metadata (bluecodec_config).
# pip install "bluecodec @ git+https://github.com/maxmelichov/blue-codec.git"
from bluecodec import BlueCodec
codec = BlueCodec.from_pretrained("notmax123/blue-codec")
latents = codec.encode(audio) # audio [B, L] at 44.1 kHz -> latents [B, 24, L // 512 + 1]
audio_out = codec.decode(latents)
| input | 1253-channel spectrogram (1025 log-linear + 228 log-mel; FFT 2048, hop 512) |
| encoder | Conv1d stem + 10 ConvNeXt blocks + projection to 24 channels (~25.6M parameters) |
| decoder | causal Conv1d stem + 10 causal dilated ConvNeXt blocks + vocoder head (~25.3M parameters) |
| latent | 24 channels at 44100 / 512 = 86.13 Hz |
BlueV3 (notmax123/BlueV3, notmax123/BlueV3-onnx) was trained in this latent space: its latent statistics and voices are valid only with this codec.
Earlier files (the 1.5M-step codec and the September 2026 encoders) stay available under the git tag v3.0.0,
e.g. BlueCodec.from_pretrained("notmax123/blue-codec", revision="v3.0.0").
License
The decoder is a fine-tune of Supertone Inc.'s Supertonic-3 vocoder
(Supertone/supertonic-3, BigScience OpenRAIL-M), so its weights are a
derivative and the OpenRAIL-M use-based restrictions apply (LICENSE.OpenRAIL-M). The encoder contains none of
Supertone's weights. Code: maxmelichov/blue-codec (MIT).