Samleng សំឡេង

Natural Khmer speech, open source

Samleng is an open text-to-speech model. It reads numbers, names and Khmer–English mixed sentences correctly, and runs on an ordinary computer.

(01) Listen

Same sentence,
three models

Pick a sentence and compare the base model, Samleng, and the teacher it learned from. These come from our test set, not hand-picked best cases. The score on the right is a blind listener rating from 1 to 5.

    (02) Results

    Measured by machine,
    judged by Khmer ears

    Every model reads the 300 Khmer sentences of KhmerTTS-Bench. A speech recognizer writes down what it hears, and we count the characters it gets wrong.

    ModelKhmer error rateUnseen English wordsListener score (1–5)
    Base modelBefore our training
    10.4%
    		36.7%2.16
    Samleng v1Trained on 1.2 h of teacher audio · 36M
    5.7%
    		23.7%3.40
    Teacher · VoxCPM22B parameters · needs a GPU
    3.1%
    		12.9%3.72
    Gemini 3.8 TTSGoogle's paid service
    2.2%
    		–3.68
    • Error rate

      Median share of characters the recognizer gets wrong. Scores move about 1 point by chance, so smaller gaps are not reported as real.

    • Listener score

      Blind 1–5 ratings of 25 sentences. So far there is one rater (the author); we are recruiting more native listeners.

    • Unseen English words

      Khmer sentences with English words the model never saw in training. Gemini reads raw text; the other models use our number and date front end.

    (03) How it works

    A big teacher,
    a small student

    Large models speak Khmer well, but they are slow and need a GPU. We use one as a teacher to train a small model that runs fast on an ordinary computer.

    1. ខ្ញុំរីករាយណាស់ដែលបានជួបអ្នក។

      9,042sentences read by the teacher

      The teacher reads

      VoxCPM2, a 2-billion-parameter model, reads every sentence in one fixed voice.

    2. One dot = 100 clipsrejected

      574clips thrown out

      Every clip is checked

      A speech recognizer and a native ear reject clips that are misread, off-voice or too fast. 8,468 clips (8.5 hours) are kept.

    3. Teacher · 2B
      Student · 36M

      56×smaller than the teacher

      The student learns the voice

      A 36-million-parameter model learns from the kept clips and runs faster than real time on a regular CPU.

    (04) KhmerTTS-Bench

    An open yardstick
    for Khmer speech

    320 sentences reviewed by a native speaker: pure Khmer, long sentences, names, numbers, Khmer–English mixed and English, with fixed scoring code. Anyone can score their model and compare it fairly.

    Open on Hugging Face Leaderboard
    320sentences, native-reviewed
    6sentence categories
    9systems measured
    CC BYfree to use

    (05) Lab log

    What we tried,
    including what failed

    1. The big run

      8.5 hours of teacher audio. It crashed twice because the Mac's GPU ran out of memory; we found the cause, fixed it, and resumed from the last checkpoint.

    2. First listener scores

      Blind ratings: base model 2.16 → Samleng v1 3.40. Steady speaking pace won both by ear and by the numbers.

    3. A new teacher voice

      We picked a clearer synthetic voice and had the teacher re-read every sentence.

    4. KhmerTTS-Bench published

      Measured all 7 open Khmer TTS models and two Gemini versions. Gemini 3.8 is the most accurate, but it is a paid service.

    5. The first student

      Trained on just 1.2 hours of teacher audio, it made about 43% fewer errors on Khmer sentences.

    6. Metrics can mislead

      On Khmer–English sentences, the plain error rate did not match what people hear. We built a new measure and checked it against listener scores.

    (06) Join in

    Help rate Khmer voices

    Only native speakers can tell which voice sounds truly Khmer. It takes about 15 minutes: listen to short clips and score each from 1 to 5.

    I want to help