A $3 chip with 512 KB of SRAM and no NPU can now run an entire neural text-to-speech stack. That is the result documented in the sanoTTS repository, which presents a family of TTS models ranging from 294k to 2.2m parameters, with 11 voices and 6 languages. The smallest model, in INT8 quantization, occupies 337 KB: enough to run on a low-cost microcontroller without a dedicated neural network accelerator.

The numbers show extreme compression: according to the project, sanoTTS is about 1000 times smaller than Kokoro and 9000 times smaller than Voxtral TTS; the 1.5-million-parameter variant records SCOREQ 4.13 and UTMOS 4.10, beating larger models such as Inflect Nano (4.63m) and KittenTTS (15m), which score 3.81 and 3.02 respectively. On an ESP32 microcontroller, the reported RTF is 0.225—one second of processing produces four seconds of audio.

Memory, not compute, becomes the bottleneck

The most important part is not the single benchmark but the shift in constraints. On a chip without an NPU, the real limit is available memory, not raw performance; INT8 quantization turns the stack from a capacity problem into an architecture problem. This makes voice inference a local function on hardware drawing just a few watts, with immediate advantages: no network latency, no per-character cost, no audio stream sent to cloud services. For makers of home devices, industrial panels, toys, or accessibility tools, the result changes the edge-versus-cloud equation.

A WebAssembly port extends the same principle to the browser, and the published recipe for adding languages and voices points to a modularity that needs no dedicated infrastructure. For the industry, the project signals that the race toward ever-larger models is not the only trajectory: at the edge, compression and memory efficiency shift value from API vendors to hardware makers. It is not a theoretical threat to cloud TTS services, but it shrinks their addressable space where data sovereignty, offline operation, or operating cost matter.

What remains to be assessed in the field is perceived quality on extended languages and voices: the benchmarks indicate around 2% WER measured with Whisper, but real-world adoption will depend on robustness in noisy scenarios and ease of integration. For AI-RADAR, cases like this make a usually abstract trade-off tangible: if a voice pipeline can be compressed to 337 KB, the argument for keeping speech processing in the cloud will need reasons beyond model size.