On-Device Text-to-Speech
Small text-to-speech models that run offline on phones, CPUs or microcontrollers without a cloud service.
Tier Grades
Tier A: Frontier AI labs
feasible
Frontier labs could do this easily, though the exemplars come from small teams.
Tier B: Indonesian AI labs
feasible
Kokoro was trained for about USD 1,000 of A100 time, well within a national lab budget.
Tier C: Institution units such as ITB
feasible
An 82M parameter model trained on about 1,000 GPU hours fits a university unit.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
sanoTTS already speaks Indonesian on a USD 3 microcontroller, so offline voice proofs of concept need no training.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeKokoro: 82 million parameters, about 1,000 A100 80GB GPU hours in total (about USD 1,000); sanoTTS: 294k to 2.27M parameters, real time on an ESP32-S3.
- DataKokoro card: under 100 hours of training data for v0.19 and a few hundred hours for v1.0; sanoTTS covers 11 voices across 6 languages including Indonesian.
- Team size (proxy)Proxy only: not in the sources read (model cards, no papers).
- Time horizonWeeks to months.
- Shelf-life signalsOperating systems shipping good offline voices in Indonesian
- Pipeline roleSmall teams release compact voices; Tier C and Tier B train Indonesian and regional voices; Tier D puts them on devices for offline field use.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Links
R&D TopicOn-Device Text-to-Speech
Models
sanoTTSAmpixa, speech synthesis
Kokoro-82Mhexgrad, speech synthesis
PiperRhasspy, speech synthesis
NeuTTS AirNeuphonic, speech synthesis+4 more
sanoTTSAmpixa, speech synthesisKokoro-82Mhexgrad, speech synthesisPiperRhasspy, speech synthesisNeuTTS AirNeuphonic, speech synthesis+4 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 8
Kokoro-82Mhexgrad, speech synthesisexemplarsanoTTSAmpixa, speech synthesisexemplarKokoro v0.19hexgrad, speech synthesisproduced by
MeloTTSMyShell.ai, MIT (Jameel Clinic), speech synthesisproduced byNeuTTS AirNeuphonic, speech synthesisexemplarPiperRhasspy, speech synthesisexemplar
Silero TTSSilero, speech synthesisproduced bySupertone supertonic-3speech synthesisproduced byPapers
- Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram PredictionsJonathan Shen et al., ICASSP 2018 (per arXiv comment), 2017 founding
- Matcha-TTS: A fast TTS architecture with conditional flow matchingShivam Mehta et al., ICASSP 2024 (per arXiv comment), 2023 state of the art
- StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language ModelsYinghao Aaron Li et al., NeurIPS 2023 (per arXiv comment), 2023 state of the art
- Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-SpeechJaehyeon Kim et al., ICML 2021 (per arXiv comment), 2021 state of the art
- LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture SearchRenqian Luo et al., ICASSP 21 (per arXiv comment), 2021 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- hexgrad/Kokoro-82M , Hugging Face huggingface.co https://huggingface.co/hexgrad/Kokoro-82M
- huggingface.co/ampixa/sanoTTS https://huggingface.co/ampixa/sanoTTS
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "On-Device Text-to-Speech", https://aurion.semai.id/topics/on-device-text-to-speech/, records of October 2026.