INGMUYLEANG/VoxPocket-v4

VoxPocket: small Khmer text-to-speech

VoxPocket is a 0.8B Khmer TTS model (44.1 kHz). It was fine-tuned from VoxCPM 1.5 on real Khmer speech and on synthetic speech distilled from VoxCPM2. It reads Khmer, Khmer mixed with English words, and plain English. It can clone a voice from a short clip and take a style or a voice description.

🗣️ Try it in your browser, no download needed: VoxPocket demo. It has ready voices, styles and voice design.

Use

from voxcpm.core import VoxCPM
import soundfile as sf

m = VoxCPM.from_pretrained(hf_model_id="INGMUYLEANG/VoxPocket-v4", load_denoiser=False)
sr = m.tts_model.sample_rate  # 44100

# 1) clone a voice: a clean 4–9 s clip + exactly what is said in it
wav = m.generate(text="សួស្តីបងប្អូនទាំងអស់គ្នា អរគុណច្រើនសម្រាប់ការគាំទ្រ។",
                 prompt_wav_path="my_voice.wav", prompt_text="<words spoken in my_voice.wav>",
                 cfg_value=2.0, inference_timesteps=10)
sf.write("clone.wav", wav, sr)

# 2) clone + style: put the style in brackets before the text
wav = m.generate(text="(sad, low and slow tone)សួស្តីបងប្អូនទាំងអស់គ្នា។",
                 prompt_wav_path="my_voice.wav", prompt_text="<words spoken in my_voice.wav>")

# 3) design a voice, no clip: describe it in brackets
wav = m.generate(text="(A young woman, clear and bright voice, cheerful tone)សួស្តីបងប្អូនទាំងអស់គ្នា។")

Tips:

• Reference clips: keep them short, about 4–9 s, and clean, with the exact words that are spoken.
• Long text: split it into sentences.
• Doubled words (ៗ): write them out (ភ្លាមភ្លាម) or let your text step expand them; English words next to Khmer are fine.
• Styles: the phrases used in training work best, for example sad, low and slow tone, excited and energetic, whispering quietly, happy, smiling tone, calm, slow and steady, serious and clear news-reader tone.

Results

The test reads 40 held-out Khmer sentences with the same voice clip (cfg 2.0, 10 steps). A Khmer speech recognizer (seanghay/Qwen3-ASR-0.6B-Khmer) listens back, and we count letter errors (lower is better):

	median	mean	perfect sentences	sentences with > 10% errors
VoxPocket	5.4%	6.9%	8 / 40	11 / 40

Training

• Starting point: VoxCPM 1.5, fine-tuned on Khmer over several rounds.
• Data: about 670 h in the training list.
  • 237 h of real Khmer speech
  • 270 h of synthetic VoxCPM2 speech, all checked by a speech recognizer: Khmer, mixed, English, clone, style and designed voices
  • 79 h of voice clip + style pairs, so a cloned voice can change its tone
  • 5 h of English → Khmer switch sentences (doubled words and Khmer endings right after English)
• Part of the synthetic speech is published as INGMUYLEANG/voxpocket-khmer-tts-v4 (gated).
• Hardware and time: see Training hardware below.

Training hardware

GPU	GPU memory	Training run
1× NVIDIA GeForce RTX 5090	32 GB (CUDA 13)	newest run: about 1,010 h of audio, about 15 h

Limits

• Styles on cloned voices are still weak: the tone changes only a little. Designed voices (no clip) follow descriptions better.
• Rare words and names are sometimes swapped for common ones. English loanwords inside Khmer can sound Khmer-ish.
• English is basic. The model is made for Khmer first.
• Do not use it to impersonate real people or to make deceptive audio. Only clone voices you have permission to use.

Citation

If you use this work, please cite:

@misc{ing2026voxpocket,
  author    = {Ing, Muyleang},
  title     = {VoxPocket: A Small Khmer Text-to-Speech Model with Voice Cloning and Styles},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/INGMUYLEANG/VoxPocket-v4}
}

The dataset:

@misc{ing2026voxpocketdata,
  author    = {Ing, Muyleang},
  title     = {VoxPocket Khmer TTS: A VoxCPM2 Distillation Dataset for Khmer Speech},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/INGMUYLEANG/voxpocket-khmer-tts-v4}
}

VoxPocket is built on VoxCPM 1.5 and its synthetic data was made with VoxCPM2 (OpenBMB); please also credit VoxCPM when you use it.

Author

Ing Muyleang, built with the help of AI tools (agy, Claude).

Contact

Questions: contact the author through the Hugging Face profile or GitHub.