A new open-source speech synthesis model squeezes convincing audio generation into 264,800 parameters, small enough to run on a microcontroller instead of a server rack.
Researchers built GrainSpeech after testing how much surrounding context a compact voice model actually needs. They found that letting the encoder's self-attention look beyond 15 phonemes added no consistent improvement to pitch, energy, or duration accuracy, so they swapped in a fixed-receptive-field convolutional encoder instead. That change cut prediction errors by 36.0% for pitch, 17.3% for energy, and 3.4% for duration. The team also adapted a gradient-variance technique borrowed from image processing, reworking it for Mel-spectrograms with axis-specific gradients and log-domain variance matching, to recover fine detail that a plain transfer had degraded. The result runs 17.9 times faster than real-time on microcontroller hardware and scores comparably to far larger text-to-speech models on the UTMOS quality metric, using less than 1.5% of their parameter count.
That last part is the real story in context: text-to-speech systems have been trending toward billions of parameters running on GPUs in someone else's data center. A synthesizer small enough to fit in a microcontroller's memory can run on a wearable, a hearing aid, or an offline IoT device, with no network round-trip and no per-request cloud bill, in exchange for the flexibility a large model brings.
UTMOS is a machine-predicted proxy for how humans would rate voice quality, not a listening panel, so treat "comparable to models 65-plus times its size" as promising rather than settled until independent ears weigh in.