A Khmer speech-to-text (STT) model — openai/whisper-base fine-tuned on Khmer speech. Built for iAny, the offline, on-device Khmer AI platform, and released open source so anyone can build Khmer voice tools.
Runs fully on-device / offline in the iAny app and mobile app. Part of iAny's mission: the best open Khmer AI, with the community and for the community.
What it does
Transcribes spoken Khmer audio (16 kHz mono) into Khmer text. Multilingual Whisper already knows Khmer; this fine-tune sharpens it on real Khmer corpora + community phone/room audio.
1import{ pipeline }from'@huggingface/transformers'2const asr =awaitpipeline('automatic-speech-recognition','sengtha/whisper-base-khmer',{dtype:'fp32'})3const{ text }=awaitasr(audioFloat32Array,{language:'km',task:'transcribe'})
whisper.cpp (GGML, on mobile/edge)
./whisper-cli -m ggml-base-khmer-q5_1.bin -l km -f audio.wav
sengtha/iany-khmer-voice — real phone/room audio contributed at iany.app/voice (CC-BY-SA-4.0), oversampled so the model weights real-world conditions. This set grows with community contributions, and the model is periodically retrained.
Evaluation
Character Error Rate (CER) is the meaningful metric for Khmer; report space-normalized CER since Khmer has no word spaces and Whisper inserts them. WER is not meaningful for Khmer (ambiguous word segmentation).
Test set
CER (space-normalized)
CER (raw)
FLEURS km_kh test (standard benchmark, out-of-domain read speech)
~22.8
~23.9
In-domain dev (training corpora)
~19
—
FLEURS is clean, formal read speech with foreign proper nouns, so it's a harder, out-of-domain stress test — treat it as a comparability anchor, not a ceiling.
Limitations
Best on clear, careful speech; spontaneous/noisy phone audio in a room is harder for a base-size model. More community /voice data + retraining is the path to better real-world accuracy.
Weak on numbers and embedded English/proper nouns.
On low-confidence audio it can emit byte-fallback characters (rare); downstream apps may strip U+FFFD.
Intended use & responsible use
Voice input, dictation aids, transcription, and Khmer accessibility tools. It is an aid, not a certified transcription service — review output where accuracy matters.
License & attribution
Released under CC-BY-SA-4.0, inherited from the DDD-Cambodia training data. You must credit DDD-Cambodia and share derivatives alike. The base model openai/whisper-base is MIT.
