VoxPocket reads Khmer and Khmer mixed with English aloud in a natural voice. Clone a voice from a 4–9 second clip, give it a style (happy, sad, whisper…), or design a new voice from a description.
VoxPocket v5 · cfg 2.0 · 10 steps · about 6–7 s on the free demo
Try it
Type it, hear it
Type Khmer (or Khmer with Latin letters), pick a voice and a mood, and listen. Free to try.
Model
Higher quality voice: powered by VoxPocket on Hugging Face. Usually ready in a few seconds; when many people use it, requests wait in line.
Text
Tap a result to add it to the text. Sing Khmer by Akkhara · model
Voice
Use a clean 4–9 s clip of one speaker. Do not copy real people's voices to trick anyone.
Style
Your clips
Listen
14 voices, many moods
Clips from the open preview dataset: one per voice, with a style. Some mix English words into Khmer, the way people really talk.
← swipe to hear all 14 voices →
WhisperKhmer
ប៉ាស្ទ័រខ្លួនឯងគឺមិនមានការភ័យខ្លាចឡើយ។
soklin · woman
ExcitedKhmer
លោកគឺជាមេដឹកនាំយោធានិងនយោបាយរបស់បារាំង។
setha · man
CalmKhmer + English
ចុះ ខ្ញុំ Copy Storage ផ្ញើទៅក្នុង Group ហើយ។
sophon · man
WhisperKhmer
តើរដ្ឋាភិបាលខ្មែរលោកជ្រាបទេ។
chanda · man
SoftKhmer + English
ចុះ ក្រុមហ៊ុនកំពុងរក Developer ដែលចេះ Flutter។
mey · woman
ExcitedKhmer
កោះរ៉ុងត្រូវបានប្រកាសឈប់នៅលើផ្លូវលំចេក។
oudom · man
SoftKhmer + English
ចុះ ចូលទៅ Battery រួចចុចលើ Charger។
maly · woman
SeriousKhmer
ពន្ធនាំចូលគ្របដណ្តប់លើទំនិញជាច្រើនប្រភេទ។
phanin · woman
SadKhmer + English
ពួកគេមិនបាច់ Book Room ទេ។
aura · man
SoftKhmer
ឈ្មោះខែទីបួនប្រើតាមចន្ទគតិ។
rithy · man
HappyKhmer + English
អូនសូមផ្ញើ Invoice ឆាប់ឆាប់។
neary · woman
CalmKhmer
រាត្រីថ្ងៃនោះខ្ញុំព្រះករុណាសិងមិនលក់ពេញមួយយប់។
seth · man
SoftKhmer + English
ការប្រមូលវាយនភ័ណ្ឌ LACMA។
bora · man
SeriousKhmer
មានប្រជាប្រិយភាពខ្លាំងក្លានាសម័យនោះ។
theary · woman
Use
Three ways to make speech
1 · Clone a voice
Give a clean 4–9 s clip and exactly what is said in it. VoxPocket reads new sentences in that voice.
2 · Clone + style
Put a style in brackets before the text: (sad, low and slow tone)សួស្តី… Happy, sad, angry, whisper, surprised, news reader…
3 · Design a voice
No clip at all: describe the voice. (A young woman, clear and bright voice)សួស្តី…
from voxcpm.core import VoxCPM
import soundfile as sf
m = VoxCPM.from_pretrained(hf_model_id="INGMUYLEANG/VoxPocket-v5", 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។", prompt_wav_path="my_voice.wav",
prompt_text="<words spoken in my_voice.wav>", cfg_value=2.0, inference_timesteps=10)
# 2) clone + style
wav = m.generate(text="(angry and upset tone)ហេតុអ្វីបានជាអ្នកមកយឺតម្ដងទៀត?", prompt_wav_path="my_voice.wav",
prompt_text="<words spoken in my_voice.wav>")
# 3) design a voice, no clip
wav = m.generate(text="(A young woman, clear and bright voice, cheerful tone)សួស្តី។")
sf.write("voxpocket.wav", wav, sr)
Tips: short clean reference clips with their exact words; split long text into sentences.
Results
Reads Khmer better than its 2.4B teacher
40 held-out Khmer sentences, same voice clip, cfg 2.0, 10 steps, seed 0. A Khmer speech recognizer (seanghay/Qwen3-ASR-0.6B-Khmer) listens back and counts letter errors. Lower is better.
4.3%median letter error (VoxCPM2: 4.8%)
2.7×smaller than VoxCPM2
3.6×faster than real time
64 msto the first sound (streaming)
40 held-out sentences
median error
mean error
sentences > 10% errors
VoxPocket v5 (0.89B)
4.3%
6.1%
8 / 40
VoxPocket v4
6.6%
7.2%
10 / 40
VoxCPM2 (2.38B, the teacher)
4.8%
7.0%
12 / 40
Speed, memory and power measured on 1× NVIDIA GeForce RTX 5090 (the GPU that trained v5) with the same settings for all models: VoxPocket 3.8 / 6.3 GB peak GPU memory (short / long text) vs 6.5 / 8.9 GB, 0.80 vs 1.29 Wh per minute of speech. On a second machine the same quality test gives 3.7% median error (v4: 5.1%).
Docs
How VoxPocket works
VoxPocket is VoxCPM 1.5 (an open, tokenizer-free TTS model) fine-tuned on Khmer over several rounds, with speech distilled from the bigger VoxCPM2.
The data · ~1,010 h
237 h of real Khmer speech (double weight, so voices stay natural), ~600 h of synthetic VoxCPM2 speech checked by a speech recognizer (kept only ≤ 15% errors), and 157 h of voice clip + style pairs. 340 h are new in v5: emotions, long reading, English words inside Khmer.
The model · 0.89B
A text-semantic language model (24 layers) plans the speech; a residual acoustic model and a local diffusion transformer add the voice and tone; the AudioVAE decoder makes the 44.1 kHz waveform. No audio tokenizer.
The training
From the previous VoxPocket, one more full pass over the data (778,028 clips): learning rate 2e-5 with warm-up, 43,200 steps, about 14 h on one RTX 5090. Every release is checked by listening back with a Khmer speech recognizer.
Limits. Very long text in one go still makes mistakes: split it into sentences. Styles change cloned voices less than designed voices. Rare names can be swapped; English is basic. Excited, Angry and Surprised are used with Khmer text for now; for English-only text they are coming soon (the demo reads those lines in a plain tone). All styles work with Khmer and Khmer + English text. Do not impersonate real people; clone only voices you have permission to use.
Dataset
VoxPocket Khmer TTS dataset
Synthetic Khmer speech made with VoxCPM2, every clip read back by a speech recognizer and kept only when it was right.
@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-v5}
}
VoxPocket is built on VoxCPM 1.5 and its synthetic data was made with VoxCPM2 (OpenBMB); please also credit VoxCPM. Model license: Apache-2.0.
