MERaLiON OmniVoice Hokkien TTS

This model adapts OmniVoice for Singapore Hokkien speech generation. A*STAR fine-tuned the upstream model with supervised training and GRPO. Chinese-character input is intended to produce Hokkien pronunciation, including text with English or Malay loanwords.

A regional speech generator

The published architecture is OmniVoice. This is a trained adaptation rather than a hosted copy of the parent. It is separate from MERaLiON AudioLLM models, whose generated output is text.MERaLiON OmniVoice Hokkien TTS published configurationMERaLiON OmniVoice Hokkien TTS official model card

A limited published comparison

The card evaluates 25 Hokkien sentences across everyday topics and several quality measures. This small set is useful evidence of the intended language task. It cannot establish performance across all speakers, dialect variation or long passages.MERaLiON OmniVoice Hokkien TTS official model card

Weights and TTS-specific license copy

The public repository supplies weights and a 2026 custom license that explicitly lists the Hokkien model. It includes notice and acknowledgement duties and named upstream terms. Do not infer a plain Apache or MIT license from the underlying implementation.MERaLiON Hokkien TTS repository license, 2026

Review pronunciation as well as voice quality

Practical guidance. Have a Hokkien speaker review the intended text and pronunciation. A smooth voice can still read a word incorrectly. Keep the reference transcript aligned with its audio, and compare neutral generation with the chosen reference voice before using longer passages.MERaLiON OmniVoice Hokkien TTS published configuration

Specifications & access

MERaLiON OmniVoice Hokkien TTS specifications
DetailValue and conditions
AccessDownloadable
Model identifiersMERaLiON-OmniVoice-Hokkien-TTSMERaLiON-TTS
Input and outputChinese-character text to 24 kHz mono WAV audio.MERaLiON OmniVoice Hokkien TTS official model card
Voice prompt guidanceA clean 3–8 second reference clip with matching transcript; 16 kHz mono is recommended.MERaLiON OmniVoice Hokkien TTS official model card
Base implementationOmniVoice, linked to the k2-fsa release by the official card.MERaLiON OmniVoice Hokkien TTS official model card

Related models

  • A third-generation speech-text model with regional fine-tuning for audio questions, dialogue summaries and contextual analysis.

    Launch date not confirmedDownloadable