Aloud GGUF models

These are self-contained GGUF v3 files for Aloud, a Rust-native Audio8 ASR and TTS inference engine. Aloud reads GGUF as a container and executes the models with its own Rust CPU and wgpu kernels. It does not use ggml, llama.cpp, C++, ONNX, PyTorch, or another inference runtime.

The repository has mixed per-file licensing. Read the table and the retained files under licenses/ before downloading or redistributing a model. In particular, the ASR files are non-commercial and the TTS-0.1b files have a commercial-use revenue cap.

Files

File	Architecture	Quantization	SHA-256	Upstream and license
audio8-asr-0.1b-q8_0.gguf	audio8_asr	Q8_0	6d69e49074eb48ab251bfd0b37f9b510d0cf3111f137731bed69adf442e2e9f1	Audio8/Audio8-ASR-0.1B, CC BY-NC 4.0
audio8-asr-0.1b-f32.gguf	audio8_asr	f32	f6a77447659639ee1a4aa45ff4cd280ed6787284e3db07d9e451f32f5a8ce90c	same as above
audio8-tts-0.1b-q8_0.gguf	audio8_tts_01	Q8_0	de83d5a15e57638cbaf3bf3bcb1a8646951bba10a306b6a1375d6e63c0b6b661	Audio8/Audio8-TTS-Preview-0.1b, Audio8 Community License v1.0
audio8-tts-0.1b-f32.gguf	audio8_tts_01	f32	99ccf9402b4f2622ac536f56f757d07c9239e7cfb2bd8664b5d7318788fca241	same as above
audio8-tts-0.6b-q8_0.gguf	audio8_tts_06	Q8_0	02755046ebf5cf3b7aa8258456676eab7d884dc833dba8897955b0f269ae778b	Audio8/Audio8-TTS-Preview-0.6b, Apache-2.0
audio8-tts-0.6b-f32.gguf	audio8_tts_06	f32	7861934b565ea655e34513b68e8bf0be2eed22f32b6adbf95bf3a46e44c4a9cb	same as above

The neural audio codec embedded in every TTS file comes from the pinned TTS-0.6b repository above. Its source codec.pth SHA-256 is c310505aa11fe2f6cc63b8d3130dc7e77e73227774f5c62575769b1f47a8d048. The retained Apache NOTICE records that the model architecture is inspired by the DualAR design used in Fish Audio S2 Pro.

The source checkpoint hashes verified before conversion are:

• ASR model.safetensors: 971d17f64ef5f193fca567fa3e9dc063c4eede97faabb11c0c6abf0b92b23ca4.
• TTS-0.1b model.safetensors: 69b162eb71da66b3a3f6ed7f9aa5e623853526717fd798d347b34919d58a4734.
• TTS-0.6b model.safetensors: 62dcff0adf6c2535b3260467a7c1d482b556da57266c96a444518b76e140d2c3.

Self-contained TTS

Each TTS GGUF contains its backbone, tokenizer, configuration, codec decoder, codec encoder, and stable embedded voices named sky and aiden. The voice conditioning grids were made from short synthetic reference clips for Aloud. The reference audio is not included. Normal synthesis does not load codec-encoder tensors into resident memory; Aloud selects them only while creating an external .aloudvoice file.

Usage

aloud transcribe recording.wav \
  --model hf://isala404/aloud/audio8-asr-0.1b-q8_0.gguf

aloud tts "Hello from Aloud." -o hello.wav \
  --model hf://isala404/aloud/audio8-tts-0.1b-q8_0.gguf \
  --voice sky

aloud voice create reference.wav \
  --model hf://isala404/aloud/audio8-tts-0.1b-q8_0.gguf \
  --text "Exact words spoken" \
  -o mine.aloudvoice

aloud tts "This uses my external voice." -o mine.wav \
  --model hf://isala404/aloud/audio8-tts-0.1b-q8_0.gguf \
  --voice mine.aloudvoice

Aloud downloads hf:// sources through the maintained Rust hf-hub client into the standard Hugging Face cache. It reuses cached files, works offline when the file is cached, honors standard cache and authentication environment variables, and never invokes the hf CLI during inference.

Verification

The published bytes are release outputs, not renamed upstream checkpoints. Conversion folds codec weight normalization, writes Aloud's optimized tensor layouts, preserves native GGUF Q8_0 blocks, aligns tensor payloads to 64 bytes, embeds metadata and voices, and emits per-tensor CRC plus whole-file SHA-256 checks.

Release verification covers PyTorch stage parity, ASR transcripts, exact f32 TTS token grids, Q8 teacher-forced grids, codec waveform SNR, embedded and external voices, Rust voice creation, CPU/wgpu kernels, library and CLI APIs, and TTS-to-ASR round trips. Golden tensors, source checkpoints, temporary voice audio, and test outputs are deliberately not published here.

On an Apple M3 Pro, the release Q8 ASR processes 2.66 seconds of audio in 0.171 seconds versus the pre-GGUF 0.188-second baseline. Q8 TTS-0.1b produces 2.83 seconds of audio in 1.024 seconds versus the 1.492-second baseline. Resident model storage is 329.2 MiB for ASR, 370.1 MiB for TTS-0.1b inference versus 418 MiB previously, and 855.5 MiB for TTS-0.6b inference versus 918 MiB previously. Voice creation maps 210.8 MiB of encoder-specific tensors instead of loading the TTS backbone and decoder.

After publication, all six GGUF files were downloaded into an empty cache and matched the SHA-256 values above. Real hf://isala404/aloud/... smoke tests passed for ASR, TTS-0.1b with sky, TTS-0.6b with aiden, Rust external-voice creation, external-voice TTS, cache reuse in offline mode, and exact TTS-to-ASR transcripts.

Licenses

• licenses/Audio8-ASR-0.1B-CC-BY-NC-4.0.txt applies to both ASR GGUF files. Attribution: Audio8, “Audio8-ASR-0.1B.” Commercial use is not granted.
• licenses/Audio8-TTS-Preview-0.1b-LICENSE.md applies to both TTS-0.1b GGUF files. Attribution: Copyright © 2026 Audio8. Commercial use under that license is limited to entities with annual revenue below US$2,000,000; larger entities need a separate written license.
• licenses/Audio8-TTS-Preview-0.6b-Apache-2.0.txt and licenses/Audio8-TTS-Preview-0.6b-NOTICE apply to both TTS-0.6b GGUF files and to the shared codec included in both TTS families.

No Audio8 trademark rights or endorsement are implied.