so there's no flash<br>of the wrong palette before React hydrates. Mirrors the useTheme hook<br>(storage key "rw-theme"; classes rw-theme-dark|light|ascii). -->

Owen Song releases complete local speech synthesis in 9.36M parameters - RuntimeWire

RuntimeWire

You're browsing RuntimeWire with JavaScript disabled. Articles and<br>navigation work fully. Interactive features — search, comments,<br>and newsletter signup — require JavaScript.

Why it matters

Inflect packages the complete neural speech path into 37.53 MB, giving local applications a credible voice option while exposing the feature and reproducibility costs of extreme compression.

On July 24, 2026, Owen Song released Inflect-Micro-v2, a 9,356,513-parameter text-to-speech model that takes English text through to a 24 kHz waveform without a hosted service or separate neural vocoder.

Song describes himself as a solo developer focused on machine learning efficiency. The model card says he built and funded Inflect v2 independently, and no company or outside investors are attached to the project.

That constraint shaped the release. With v2, Song was testing how far a complete tiny TTS stack could be pushed while still producing usable speech.

A complete model inside the parameter count

Small speech-model comparisons can hide part of the deployment burden by reporting an acoustic model's size while leaving out a separate vocoder. Inflect's 9.36 million figure covers the deployed neural path: text encoding, duration prediction, latent speech generation and the integrated waveform decoder. The English normalization and eSpeak-ng phoneme frontend still run as conventional CPU-side code.

The open-source repository describes Inflect as a VITS-family system with a transformer text encoder, stochastic duration prediction, residual coupling flows and an alias-reduced decoder. Song released the code and weights under Apache 2.0.

Inflect v2 comes in two sizes through the same Python interface. Micro uses 37.53 MB of FP32 weights and prioritizes output quality. Inflect-Nano-v2 contains 3,966,721 parameters in a 15.97 MB FP32 package and prioritizes footprint. Both support CPU or CUDA inference, deterministic seeds, speed and delivery-variation controls, and punctuation-aware splitting for longer passages.

Song also published a separate verified ONNX export for CPU, CUDA and DirectML. The ONNX package keeps eSpeak-ng and the Python frontend outside its two neural graphs, which total about 37.75 MB. No integer-quantized release is available; Song's documentation says basic weight quantization can audibly damage the waveform decoder.

Developers can test Micro and Nano in the Inflect v2 playground. The reference runtime is non-streaming, so it returns a complete chunk after synthesis rather than producing audio incrementally.

The benchmark is promising and appropriately limited

Inflect's performance figures come from Song's own evaluation package. Micro recorded a 66.2% preference rate in an anonymous community listening test consisting of 21 wins, 10 losses and three ties. That is 34 comparisons, and Song labels the result descriptive community evidence rather than a formal mean-opinion-score study.

The model card reports a 4.395 UTMOS22 score, with a 95% bootstrap confidence interval from 4.381 to 4.408. UTMOS22 is a learned naturalness predictor, not a human listening score. Micro also posted a 3.99% headline word-error rate derived from the equal-weight average of Qwen3-ASR and Nemotron 3.5 evaluations. Its separate Whisper large-v3 result was 2.73%.

On a managed Hugging Face instance with eight virtual CPUs and four framework threads, Micro generated audio at 6.28x real time. Nano reached 10.72x. Those results support Song's CPU-capable claim, though his own comparison shows faster compact systems: Piper Low reached 31.37x and KittenTTS Nano reached 13.33x under the directional test.

Song does not present Inflect as the speed leader. The comparator run also used a shorter 50-prompt test while Inflect's headline runtime came from 100 prompts, and some rivals ran through optimized ONNX implementations while Inflect used its canonical PyTorch runtime. The documentation keeps those differences visible instead of flattening the results into a single ranking.

KittenTTS Nano, one of Song's chosen baselines, contains 15 million parameters and offers eight voices. Inflect Micro is smaller and carries a narrower feature set. Its advantage is a compact, self-contained synthesis path rather than breadth.

Song kept the scope narrow

Inflect produces one synthetic English male voice. It does not offer voice cloning, selectable speakers, multilingual output or acoustic streaming. Numbers, abbreviations, uncommon names and homographs can still cause pronunciation errors, while independently synthesized chunks can make long passages sound less coherent.

The model has not been validated for medical, legal, emergency or accessibility-critical communication. Song originally...