Enabling Quechua speech in humanoid social robots through fine-tuned neural TTS on the UBTech Walker Tienkung

Humanoid social robots deployed in multilingual societies-particularly in indigenous communities-face a critical linguistic gap: their speech interfaces are predominantly designed for high-resource languages, excluding hundreds of millions of indigenous language speakers from equitable human-robot interaction. This study presents the first neural text-to-speech (TTS) system for Southern Quechua-the most extensive indigenous language family of the Americas, with more than ten million active speakers-deployed on a humanoid social robot. XTTS-v2, a multilingual neural TTS model, was fine-tuned on 22,727 validated utterances (30.5 h, 77 speakers) of the Cusco-Puno dialect, located in southern Peru, through a tokenizer adaptation that leverages the phonological overlap between Quechua and Spanish. Fine-tuning over 15 epochs on an NVIDIA A100 40 GB GPU achieved a 35.0% reduction in training cross-entropy loss (4.084 → 2.652), confirming effective adaptation to Quechua phonology and morphology with stable training dynamics across 35,000 gradient update steps. Inference latency was evaluated on two hardware environments: the training GPU (mean latency: 2,037 ms; Real-Time Factor (RTF) = 0.472) and the embedded SoC of the UBTech Walker Tienkung humanoid robot (Jetson AGX Orin 64 GB; mean latency: 5,913 ms; RTF = 0.961). Both configurations achieved a mean RTF below 1.0, demonstrating real-time Quechua speech synthesis directly on the robot's hardware without external computing infrastructure. Intelligibility evaluation using Whisper large-v3 with automatic language detection yielded a Word Error Rate of 147.4% and a Character Error Rate of 69.1%; these values reflect the absence of a native Quechua automatic speech recognition (ASR) model rather than the quality of TTS synthesis. These results establish the first quantitative benchmark for TTS in American indigenous languages on a humanoid robotic platform and provide a reproducible framework-encompassing training configuration, tokenizer adaptation, and deployment pipeline-for extending neural speech synthesis to other low-resource languages of the Americas.