An Indie Developer Shipped Complete Voice AI in 9.36 Million Parameters. It Runs on a Laptop CPU.

Independent developer Owen Song released Inflect-Micro-v2, a text-to-waveform speech synthesis model that fits in 9.36 million parameters (37.5 MB), synthesizes audio 6.3 times faster than real-time on a standard 8-core CPU, and beat larger compact TTS systems with a 66% blind preference rate in community testing. The Apache-licensed release, built and funded by one person, demonstrates how far model miniaturization has pushed: complete, offline voice generation that needs no spectrogram intermediate, no GPU, and no cloud.

One Developer, 37.5 Megabytes, a Complete Voice Model

Frontier labs are pouring billions into trillion-parameter models. Owen Song, working alone, just shipped a complete text-to-waveform speech synthesis system in 9.36 million parameters. It synthesizes audio 6.28 times faster than real-time on a standard 8-core CPU, and the entire model weighs 37.53 MB. 1

Inflect-Micro-v2 generates audio waveforms directly from text. Most compact TTS systems split the job in two: an acoustic model predicts a spectrogram, then a separate vocoder converts it into playable sound. Inflect collapses both stages into a single end-to-end architecture, eliminating the spectrogram intermediate and the extra model it requires. That is how the complete footprint, waveform decoder included, stays under 38 MB at FP32 precision. 1

Song built and funded the project independently and released it under an Apache 2.0 license. The model handles fixed-voice English synthesis with deterministic seeds and 24 kHz mono output, and ships with both CPU and CUDA inference paths. 1

In a blind community preference test against three established compact or local TTS baselines with larger deployable weight footprints (KittenTTS Nano, Piper Low, and Supertonic 3), Inflect-Micro-v2 won 66.2% of paired comparisons. The study randomized sample order, counted ties as half-wins, and recorded 21 wins, 10 losses, and 3 ties across 34 samples. 1

The model card is candid about what those numbers do and don't prove. The preference study is described as descriptive community evidence, not formal mean-opinion scoring. The naturalness score, 4.395 on the UTMOS22 scale, comes from a learned predictor, not human listeners. The intelligibility metric, a 3.99% word error rate averaged across two ASR models, deliberately excludes a third transcription model from its headline figure because that model produced insertion-heavy errors on certain clips. The speed comparison notes that several competing systems ran on optimized ONNX runtimes while Inflect's published numbers use the standard PyTorch runtime. 1

The evaluation protocol behind those results is granular: 400 unseen English prompts per system, a published SHA-256 hash for the test corpus, exact thread counts, and the statement that no single metric captures TTS quality. 1 Few model cards from well-funded labs document their methodology this thoroughly. Inflect puts the caveats in the same table as the headline numbers.

For developers who need voice synthesis, the deployment story is simple. Download 37.53 MB of weights, run them locally on a CPU, and generate speech faster than it plays back. No API key, no per-character pricing, no network round-trip, no text leaving the device. For applications on edge hardware, in offline environments, or in contexts where sending user text to a third-party server creates a privacy liability, the infrastructure requirement drops from a cloud contract to a file on disk. The model's deterministic seeds also mean identical inputs produce identical audio, a property that matters for testing and reproducibility. 1

Inflect also ships in a smaller variant, Inflect-Nano-v2, at 3.97 million parameters and 15.97 MB, which synthesizes 10.72 times faster than real-time on the same hardware. Both models use the same end-to-end architecture and the same API. The trade-off between them is quality, not capability. 1

Song says that if the release finds an audience, he would like to continue with a broader v3 that might add more languages, voices, and stability improvements. 1

The frontier-scale story in AI is real. Trillion-parameter models accomplish tasks that 9 million parameters cannot approach. But for the growing set of developers who need specific capabilities running on their own hardware, compression is closing the gap from a direction the scaling narrative tends to miss. A single developer, an Apache license, and a 37.53 MB file just made production-grade voice synthesis possible without a data center, a GPU, or a cloud bill. For the developers who download it, the question isn't whether local voice AI works. It's what they ship next.