A lab for the curious and the complex.

Training speed

35 GFLOPS

Four cores, full precision, forward and backward passes. Our plan had guessed 30.

Two cores

~90%

of four-core speed, so Wonder can learn on two cores while it talks on the other two.

Learn your voice

~50 min

for 30 minutes of your speech, worked out from the measured speed. The real run is next.

Power while training

~7 W

Average across the board’s power rails with all four cores busy.

Thinking, on the same board

Qwen3 running in llama.cpp with Arm’s KleidiAI kernels. Writing a reply is limited by memory, not maths, so one core already gets most of the speed and the rest are free to listen and speak. KleidiAI reads prompts 1.5–1.9× faster than llama.cpp’s reference build at Q8_0.

Listening, on the same board

Moonshine speech recognition on 20 LibriSpeech clips, 2 minutes 13 seconds of speech. Every model keeps up with live speech with room to spare, and the streaming ones transcribe while you’re still talking.

Moonshine English models (MIT) on LibriSpeech test-clean (CC BY 4.0), 10 speakers. Words wrong is word error rate after lower-casing and removing punctuation. The clock held 2.4 GHz for this test.

Speaking, on the same board

Existing open voices on one core, first on their own and then while Qwen3 writes a reply on two other cores, which is how they run in a real conversation. Our bar is under 0.3 seconds of work per second of speech while thinking. Only one clears it today, and that’s the gap Wonder’s own voice has to close.

Three test sentences, median of three runs each; Piper’s first sound waits for a whole sentence, so a long one with numbers took 0.7 s alone and 1.1 s while thinking. Kokoro renders each whole sentence before any sound comes out, so it is the slowest to start.

Decoding, on the same board

A voice model’s last step turns its plan for the sound into a waveform. Mimi, the decoder our first voice was going to build on, needs 117 ms for every 80 ms of audio, so we raced the open alternatives on one core. The budget is under 80 ms, shared with the rest of the voice. Vocos does it in about 30 ms, four times faster than Mimi, and stays close to the original recording.

Each streaming call decodes 80 ms of new audio plus 240 ms of context, except Mimi, which keeps its own streaming state. Quality comes from 20 LibriTTS-R clips (24 kHz, CC BY 4.0) rebuilt from their own mel frames or tokens, scored by UTMOS22 predicted opinion and Moonshine word error rate; with 20 clips, a point of words wrong is a handful of words. The HiFi-GAN weights learned one speaker, which costs it quality here. These pretrained decoders are only measured: Wonder’s own will be retrained on openly licensed audio.

Raspberry Pi 5, 8 GB · Ubuntu 26.04 · PyTorch 2.14.1 · llama.cpp with KleidiAI · no cooler: under sustained load it reached 84 °C and slowed to 1.5 GHz, so treat these as a floor. * Measured while heat-limited. Measured with wonder-bench, which is open source: the code and raw results are public, so you can run it on your own Pi.