sombochea/zio-tts

Zio TTS — Khmer text-to-speech

A small VITS model that speaks Khmer at 24 kHz, with several voices, ONNX export, and an easy fine-tuning path. Part of the open Zio project.

Research release. Understandable Khmer with some mispronunciations; not studio quality. Khmer only, neutral voices, no published evaluation yet. Listen to samples/, which were generated by this exact checkpoint.

Use

git clone https://github.com/sombochea/zio-tts && cd zio-tts && uv sync --extra tts    # Python 3.12+, ffmpeg
hf download sombochea/zio-tts --local-dir zio-tts
uv run zio speak "សួស្តី តើអ្នកសុខសប្បាយជាទេ" hello.wav --run zio-tts

Option
--voice km_7548__neutral	voice id <speaker>__<emotion>; all ids are in vocab.json → voices (default km_6753__neutral)
--speed 1.2	faster (>1) or slower (<1)
--expressive 0.8	livelier prosody; 0.4 = steadier (default 0.667)

Python: from zio.tts.infer import synthesize; synthesize("zio-tts", "ថ្ងៃនេះអាកាសធាតុល្អណាស់", "out.wav", voice="km_6753__neutral")

ONNX (no PyTorch): python examples/onnx_infer.py zio-tts.onnx vocab.json "សួស្តី" out.wav km_6753__neutral. Inputs: input (int64 ids, blank_id between characters), input_lengths, scales = [expressive, 1/speed, 0.8], sid (from vocab.json → voices).

Text is normalized (digits, ៛, $ → words) and word-segmented automatically. The model was trained on word-spaced Khmer.

Fine-tune

Bring ~30 min of consented audio and start from best_model.pth. New voices get fresh embeddings, and everything else is reused. Recipe: training docs.

Details

Architecture	VITS (coqui-tts) on graphemes (Khmer characters + a–z); no phonemizer, since espeak-ng has no Khmer
Data	subsets of sombochea/zio-tts-data (OpenSLR 42, seanghay/km-speech-corpus, ...). The checkpoint note below says which subsets and steps this release used
Files	best_model.pth (weights only), config.json, speakers.pth, zio-tts.onnx, vocab.json

Limitations

• Khmer only. Some voices come from 16 kHz or noisy sources and sound duller. Long or rare words may be mispronounced; keep sentences under ~15 s.
• No emotion control yet: the __<emotion> slot only has neutral until emotional recordings are contributed.
• Numbers are read up to 999,999,999; dates, units and mixed Khmer/English text are not handled.
• Don't use it to impersonate anyone. Disclose that the speech is synthetic.

Evaluation

None yet (no MOS, no CER). Treat quality claims as unverified.

License

CC BY-SA 4.0, because parts of the training data are share-alike (a precaution, not legal advice). Code: Apache-2.0. Credits: VITS (Kim et al., 2021), coqui-tts, OpenSLR 42 (Google), seanghay/km-speech-corpus, Digital Divide Data.

Checkpoint

v2, step 16,000 (~10 epochs on ~40 h: SLR42 + seanghay + DDD), initialised from v1. One Kaggle T4 session.

Sample sentences

1. សួស្តី តើអ្នកសុខសប្បាយជាទេ — samples/sample1.wav
2. ថ្ងៃនេះអាកាសធាតុល្អណាស់ — samples/sample2.wav
3. ម្ហូបនេះតម្លៃ ៥០០០៛ — samples/sample3.wav
4. កម្ពុជាជាប្រទេសមានប្រវត្តិសាស្ត្រយូរលង់ — samples/sample4.wav

Same sentence 1 in other voices:

• voice km_6753__neutral — samples/voice_km_6753__neutral.wav
• voice km_7548__neutral — samples/voice_km_7548__neutral.wav
• voice seanghay__neutral — samples/voice_seanghay__neutral.wav
• voice ddd_f-adt1-0002__neutral — samples/voice_ddd_f-adt1-0002__neutral.wav