Answers are grounded in this paper and cite their sources.
Voice-controlled humanoid robots can take over hazardous tasks in defence and field scenarios where direct human intervention is risky. Most reported prototypes couple an Android front end to an Arduino through Bluetooth, which restricts on-board signal processing, obscures end-to-end timing, and is seldom evaluated under speaker-independent conditions. This work reports the design, implementation and evaluation of an artificial neural network based speech-command recognition system that controls a humanoid robot prototype from within a single integrated development environment. A custom corpus of 5,000 utterances covering five motion commands was recorded from ten speakers (five male, five female) at 16 kHz under controlled laboratory conditions with a signal-to-noise ratio of at least 30 dB. A six-stage cleaning pipeline retained 4,728 utterances. Two speakers, one male and one female, were held out in full before any frame-level processing; the remaining eight speakers formed a development pool of 3,771 utterances that was further divided by speaker into six training and two validation speakers. Fourteen Mel-frequency cepstral coefficients were extracted from every 25 ms frame with a 10 ms hop and mean-pooled over the voiced frames of an utterance, yielding one 14-dimensional vector per utterance. A 14-50-5 feedforward network with rectified linear hidden units and a softmax output was trained with MATLAB scaled conjugate gradient backpropagation (trainscg) using speaker-defined index partitioning and validation-based early stopping. On a class-balanced speaker-independent test set of 750 utterances (150 per command) the network classified 651 utterances correctly, an overall accuracy of 86.8% with a 95% Wilson confidence interval of 0.842 to 0.890 and macro-averaged precision, recall and F1 score of 0.87. Per-command recall ranged from 84.0% for turn right to 91.3% for stop, the directional pair being hardest to separate because of a shared phonetic prefix. The classifier holds 1,005 parameters and needs about 2.0 kFLOP per utterance, some 33 and 53 times fewer parameters than the one-dimensional convolutional and long short-term memory baselines evaluated here. Command latency, measured from end-of-speech detection to actuator acknowledgement, averaged 149.7 ms with a standard deviation of 3.9 ms. The same feature extraction and classifier framework, with cleaning thresholds adapted to that corpus, was additionally applied to the public Google Speech Commands v0.02 corpus, with the released words forward, backward, right, left and stop standing in for the five commands and with the release’s own speaker-disjoint partition; on a class-balanced test set of the same size it reached 67.5% (95% Wilson interval 0.640 to 0.707) against 95.2% for the convolutional baseline, which indicates that the accuracy reported above is specific to this corpus and suggests that the loss of temporal information caused by utterance-level mean pooling may be an important limitation when the distinguishing material is short, although the present experiments do not isolate it as the sole cause. The contribution is a transparent, reproducible, speaker-independent laboratory baseline for small-vocabulary humanoid robot voice control; it is not evidence of readiness for operational deployment, which would require validation under realistic acoustic and mission conditions.
No takes yet. Share an insight, caveat, or question.
Shrivastava et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: