Speech System Steers Humanoid Robot With 86.8% Accuracy
Researchers report a speaker-independent speech command recognizer that controls a humanoid robot prototype with about 150 ms latency.
Updated on Nov. 16, 2026 in Robotics

Researchers have reported a speaker-independent speech command recognition system that controls a humanoid robot prototype. The system classified 86.8% of 750 held-out test utterances correctly.
Why it matters
The work points to lightweight voice control for robots, since the classifier holds only 1,005 parameters and needs about 2.0 kFLOP per utterance. Its accuracy on a standard public benchmark, 67.5% against 95.2% for a convolutional baseline, shows the approach still trails heavier models.
The classifier uses 1,005 parameters and roughly 2.0 kFLOP per utterance, with command latency averaging 149.7 ms (standard deviation 3.9 ms). Performance on a larger public benchmark was not established for this vocabulary.
The details
The system extracts 14 Mel-frequency cepstral coefficients from 25 ms frames and feeds them to a 14-50-5 feedforward network that recognizes five motion commands. The team recorded 5,000 utterances from ten speakers at 16 kHz, kept 4,728 after a six-stage cleaning pipeline, and fully held out two speakers before frame-level processing.
Timeline
Article published early as accepted peer-reviewed research on 2026-11-16.
The Big Picture
On the Google Speech Commands v0.02 benchmark, the lightweight network reached 67.5% accuracy while a convolutional baseline reached 95.2%, marking a size-versus-accuracy trade-off. The result frames compact speech control as a practical option for small robots rather than a replacement for larger recognizers.
Compact recognizers that need little computing power could let low-cost devices respond to spoken commands without a constant cloud connection. The accuracy gap on public benchmarks means users should expect limited command sets for now.
The takeaway
Lightweight models can handle a small set of spoken commands on modest hardware, but accuracy on broader benchmarks still lags larger networks. Builders choosing speech control should weigh vocabulary size against available compute.
Further reading
Explore more on Robotics for related developments in robot control and perception.
Source note: This article includes information reported by Nature.



