Nara Mini v1

Nara Mini v1 is a Kokoro-based text-to-speech model adapted for Burmese and English, with four preset voices and local inference support.

Nara Audio website

The four voices were developed using synthetic speech generated with VoxCPM2. The exact starting Kokoro checkpoint version has not been recorded in the available release documentation.

Voices

Voice ID	Display name
`shoon`	Shoon
`thazin`	Thazin
`ko-banyar`	Ko Banyar
`htet-myat`	Htet Myat

Each voice uses its own network and matching voicepack. Audio output is 24 kHz mono. Matching voice metadata supplies the vocabulary and style vectors.

Text processing

Inference uses Nara Mini v1's Burmese pronunciation rules, exceptions lexicon, English phonemization, and matching symbol IDs. Burmese number normalization converts supported numeric expressions to spoken words before phonemization.

Export formats

weights/<voice>/model.safetensors contains the original full-precision inference tensors. Keys have the form <module>.<parameter>, with modules bert, bert_encoder, predictor, text_encoder, and decoder. The matching architecture settings are in config.json. The vocabulary, style vectors, and length-bin centers are in onnx/<voice>/voice.json.

The mobile ONNX export in onnx/<voice>/ uses a full-precision encoder and a selectively quantized INT8 decoder, with matching voice.json metadata. packages/nara-mini-1-<voice>.nara packages the ONNX graphs and metadata with file checksums for installation in the Nara app. SHA256SUMS records release file hashes.

Run inference

Use Python 3.10 or newer.

pip install huggingface_hub
hf download YeeNyunn/nara-mini-v1 --local-dir nara-mini-v1
cd nara-mini-v1
pip install -r requirements.txt
python synthesize.py --voice shoon --text "မင်္ဂလာပါ။ Hello world." --output speech.wav

The example runs the ONNX encoder and decoder on CPU. Use --threads to set the CPU thread count. Use --no-normalize-numbers to preserve numeric text. English phonemization uses Misaki and its eSpeak dependency. Those dependencies retain their respective licenses.

Limitations

Pronunciation and prosody can vary with text and language mixing. Number normalization follows Burmese reading rules, including in mixed-language text. Long passages are processed in shorter pieces. Performance depends on device hardware, export format, and passage length.

License and acknowledgments

Nara Mini v1 is released under Apache 2.0. See LICENSE and NOTICE.

Kokoro provides the base model. VoxCPM2 provided synthetic speech used to develop the four preset voices.