Small speech models often rely on a separate neural vocoder to turn predictions into audio. This one instead predicts parameters for a compact source-filter synthesizer, keeping the complete system at about 4.5 million parameters. That makes it a practical example of how much a deliberately constrained TTS model can do on ordinary hardware—without claiming to replace larger, more expressive systems.
Key Capabilities
- CPU and browser inference: The model runs at about 70 times real time on four CPU threads, and is available in PyTorch and ONNX formats. This makes local or browser-based speech generation feasible without a dedicated GPU.
- Compact source-filter audio generation: Rather than using a neural vocoder, it shapes a predicted-pitch pulse train and noise signal with learned filters. The waveform-generation component has about 26,000 parameters, helping keep the overall model small.
- Two fixed English voices: F2 and M2 generate 24 kHz American English speech. Speed and pitch can be adjusted, but there is no voice cloning or streaming.
- Synthetic training data: The model was trained from scratch on roughly 18 hours of synthetic speech. Its reported transcription checks use only small sets, so the scores are sanity checks rather than evidence of broad benchmark performance.
Who It Suits—and the Trade-offs
A good fit for developers exploring lightweight, local English TTS, browser demos, or small-footprint speech applications. Look elsewhere if you need multiple languages, voice cloning, streaming output, robust pronunciation of unusual names, or highly natural expressive speech. The two voices are fixed, and the training recordings were themselves generated by another TTS model.