Audio8 · TTS · 0.1B
Audio8 TTS 0.1B compresses the Audio8 speech-generation approach into a compact model designed for local CPU and low-memory environments.
It targets local CPUs, low-memory devices and offline environments: a 170M-parameter main generative model with zero-shot voice conditioning, and a complete inference path built on ONNX Runtime. The value is in a usable voice-generation pipeline inside a compact budget.
Audio8 TTS 0.1B keeps the same semantic-to-acoustic handoff as the 0.6B model: Slow AR advances semantic tokens frame by frame, Fast AR reads its hidden states and fills the ten RVQ codebooks. The difference is the Slow AR backbone. It uses the Falcon-H1 hybrid design, mixing attention layers with Mamba 2 state-space layers at a reduced width of 512, while keeping the contract between long-range state and within-frame acoustic decoding.
| Component | Configuration |
|---|---|
| Main model | 170M parameters, excluding the audio decoder |
| Slow AR | 24 layers, width 512, 8 attention heads, 2 KV heads; Mamba d_state=64 |
| Fast AR | 4 layers, width 512 |
| Acoustic representation | 10 codebooks, 4,096 entries each |
| Neural audio codec | 44.1 kHz, 2,048 samples per model frame; about 120M parameters |
| Context | Up to 2,048 packed text/audio positions |
170M counts the main generative model only. The separate codec decoder still adds to storage and memory when producing a waveform.
The ONNX INT8 build is more than quantized weights. It packages the Slow AR and Fast AR token-by-token graphs, an FP16 audio codec, explicit loop state and a CPU inference interface into one path scheduled by ONNX Runtime. Once the model is downloaded, CUDA, PyTorch and Transformers are not required.
| Module | Runtime form | Role |
|---|---|---|
| Slow AR | INT8 token-by-token graph | Carries attention and Mamba state explicitly; valid-prefix attention limits computation to the valid prefix |
| Fast AR | INT8 token-by-token graph | Generates ten acoustic codebooks frame by frame |
| Audio encoder / decoder | FP16 | Registers the reference voice and reconstructs 44.1 kHz waveforms |
| Output waveform | FP32, mono | Written to a local file, a PCM stream or a service interface |
On a Linux x86_64 test machine, steady-state inference after loading uses about 0.4 GiB. On an 8-thread host, Slow AR takes about 19 ms per token and Fast AR about 8 ms per frame. Figures vary with the ONNX Runtime build, allocator, thread count and hardware.
Ordinary synthesis needs only Slow AR, Fast AR and the codec decoder. The codec encoder loads only when a new voice is created, turning a reference recording and its transcript into reusable voice codes, and can be unloaded again before synthesis resumes. Registering a voice is a one-time step, not required for every request.
From model files to the service boundary, the CLI, local HTTP, streaming PCM, voice registration and the OpenAI-compatible interface are all included.