Developer Releases 9M-Parameter TTS Model for Local Speech Synthesis

Developer Owen Song has released Inflect-Micro-v2, an open-weight text-to-speech model that fits complete speech synthesis into 9.36 million parameters. Available on Hugging Face under Apache-2.0 license, the model enables high-quality local TTS on resource-constrained hardware without cloud dependencies.

Breaking the Parameter Barrier

Most high-quality TTS systems rely on large language models or diffusion processes with hundreds of millions or billions of parameters. Inflect-Micro-v2 fits an entire synthesis pipeline—including text processing, timing prediction, and waveform decoding—into 9.36 million parameters with a 37.53 MB footprint in FP32 precision.

The model provides complete local text-to-waveform speech synthesis for English using a single fixed male voice. It runs on both CPU and CUDA with minimal memory requirements, targeting hardware that cannot support larger models.

Performance and Quality

According to the developer's evaluation documentation, Inflect-Micro-v2 achieves 6.28× real-time synthesis speed on a 4-thread CPU instance with 8 vCPUs. The model scored 4.395 on UTMOS22, a metric for predicted naturalness, and achieved 66.2% community preference in blind listening tests.

Developer benchmarks position the model as competitive with KittenTTS Nano, Piper Low, and Supertonic 3 in community preference testing. However, speed comparisons show Inflect-Micro-v2 trails Piper Low (31.37× real-time) and KittenTTS Nano (13.33× real-time) under comparable conditions.

Part of a Tiny TTS Family

Inflect-Micro-v2 belongs to the Inflect v2 series, which includes an even smaller Nano variant. Inflect-Nano-v2 contains 3.96 million parameters and occupies 15.97 MB in FP32 precision, pushing further into ultra-compact territory for embedded deployments.

Independent Development

Song built and funded Inflect v2 without company backing or outside investors. In the model's Hugging Face documentation, Song stated: "I built and funded Inflect v2 independently. If this release finds a real audience, I would like to continue the project with a broader v3."

Why It Matters

The project demonstrates feasible high-quality local speech synthesis on extremely resource-constrained hardware. By reducing parameter count to approximately 9 million, Inflect-Micro-v2 enables TTS to run efficiently on low-power CPUs and embedded devices without requiring massive VRAM or cloud connectivity. This advances tinyML for audio applications where privacy, latency, or offline operation are critical.

The release has drawn attention on Hacker News and tech coverage outlets, with observers noting the significance of complete voice synthesis at this scale. Third-party coverage has highlighted the model's potential for local-first applications, though independent benchmark replication remains pending.