nara-mini-v1

Nara Mini v1 is a Burmese and English text-to-speech model adapted by the community project Nara Audio from Kokoro. It features four voices—Shoon, Thazin, Ko Banyar, and Htet Myat—each with its own independent network, outputting 24 kHz mono audio. It includes built-in Burmese pronunciation rules, an exception word list, and number reading conventions; English uses Misaki phonetic annotation. The example script runs an ONNX encoder and an INT8 quantized decoder on CPU. The four voices were trained using synthetic speech generated by VoxCPM2.

Nara Mini v1 is a Burmese and English text-to-speech model adapted by the community project Nara Audio from Kokoro. It features four voices—Shoon, Thazin, Ko Banyar, and Htet Myat—each with its own independent network, outputting 24 kHz mono audio. It includes built-in Burmese pronunciation rules, an exception word list, and number reading conventions; English uses Misaki phonetic annotation. The example script runs an ONNX encoder and an INT8 quantized decoder on CPU. The four voices were trained using synthetic speech generated by VoxCPM2.

Why it matters Open-source text-to-speech for Burmese is rare. This model is commercially usable, includes built-in Burmese pronunciation rules and number reading conventions, and offers an ONNX export for mobile devices.

Hugging Face likes 10 · Hugging Face upvotes 10 · Hugging Face downloads 258

License: Commercial use allowed · ComfyUI: Not available · VRAM: ≈2 GB (estimated)

Reasons

  • Apache-2.0, commercially usable; attribution in LICENSE and NOTICE must be preserved upon redistribution
  • An online demo space is available on Hugging Face

Usage

  • Run hf download YeeNyunn/nara-mini-v1 --local-dir nara-mini-v1, enter the directory, run pip install -r requirements.txt, then run python synthesize.py --voice shoon --text "text to read" --output speech.wav; use --threads to set the number of CPU threads

Risks

  • The author did not record the initial Kokoro version used; pronunciation and intonation may be unstable in mixed Chinese-English text, and numbers in mixed text are read according to Burmese rules
  • Chinese is not supported; English phonetic annotation relies on eSpeak, and each dependency has its own license

Open

抱脸 text-to-speech 上升榜

Explore the latest AI Radar updates