Build1 publisher2 min readPublished
Two lines of decoder config decide whether Thonburian Whisper transcribes Thai or translates it
Mahidol's Biomedical and Data Lab publishes Thai Whisper fine-tunes that are free to use commercially, with 6.59 and 7.42 WER on Common Voice 13. Both were scored with the Deepcut tokenizer, and that shared segmentation is the reason the two numbers sit on one scale.
The Engineer · Build desk
What happened
- Thonburian Whisper is a set of Thai speech recognition models fine-tuned from OpenAI's Whisper by the Biomedical and Data Lab at Mahidol University, published free and usable commercially.
- The official Hugging Face pages for the medium and large-v3 checkpoints report word error rates of 6.59 and 7.42 on the Thai common_voice_13_0 test set.
- The sizes run from a small model the team says works on an ordinary machine up to large-v3, which needs a GPU and scores best, with distilled small and large-v3 published separately.
- Both model pages leave two sections unwritten, stating "Intended uses & limitations: More information needed" and "Training and evaluation data: More information needed".
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint A bake-off between Thonburian Whisper and a rival Thai ASR system cannot use these numbers as published; both systems would have to be rerun on one test set under one segmenter first.
- decision With commercial use permitted on both pages, the open question for a Thai transcription product is which size to run, and the write-up's suggested test is one file carrying the accents and jargon that product will actually meet.
- exposure Whoever deploys these weights absorbs the unknown on southern Thai speech, dialect and noisy recordings, because the pages do not describe how the models behave there.
- contradiction The write-up calls the medium page's training documentation complete, and also records that the same page's training and evaluation data section asks for more information, so how far the recipe is reproducible depends on which part you read.
The example code on the medium model page pins two things that Whisper otherwise guesses. It sets `pipe.model.config.forced_decoder_ids` from `get_decoder_prompt_ids(language="th", task="transcribe")`, and the page says that without those lines the model may come back in English, or translate instead of transcribing [11]. The rest of the snippet is ordinary transformers: an `automatic-speech-recognition` pipeline, `chunk_length_s=30`, and `device = 0 if torch.cuda.is_available() else "cpu"` [12].
WER is the share of words transcribed wrongly, so lower is better [4]. The gap between the two published figures is 0.83 points, about 11 percent of the higher one [23]. Both were measured on the Thai `common_voice_13_0` test set with Deepcut doing the word segmentation [6], and the write-up by Nokka is explicit that Thai WER depends heavily on how words are cut [7]. For either figure to sit next to a different Thai ASR model's WER, that model would have to be rescored on the same test set with the same segmenter. Nokka wrote that he cannot say whether the comparison holds, because each team uses a different test set and tokenizer [8].
There is a second reason not to read 7.42 as a forecast for your own audio. The medium page says Combined V4 was trained on augmented data from Thai `common_voice_13_0`, `google/fleurs` and datasets the team curated [15]. The reported evaluation is on Thai `common_voice_13_0` [6].
Both pages list `license: apache-2.0` [2]. Nokka contrasts that with Thai speech models that forbid commercial use, and concludes that an app transcribing customer audio can use these weights without asking permission [3]. He also notes that moving between sizes costs one line, the model name [13]. The small checkpoint, the size the write-up describes as running on an ordinary machine [9], has no WER on the record here [25]. Anyone shipping it on a laptop measures that themselves.
Nokka wrote that he did not run the models, and that everything in the piece comes from the Hugging Face model pages [19]. The download counts were pulled from the Hugging Face API on the day of writing [20]. The article, dated 15 September 2026, discloses that it was drafted by the deepseek-v4.1-flash model through Nous Research's Hermes Agent and then checked and edited by him [21]. The first release is credited to Atirut Boribalburephan, Zaw Htet Aung, Knot Pipatsrisawat and Titipat Achakulvisut, with the team's BibTeX giving 2022 and a DOI on Hugging Face [14]. He did not find a release date for the V4 weights [16].
What to watch
- A dated release note for the Combined V4 weights would tell users which checkpoint the 7.42 figure actually describes.
- If the team fills in the intended uses and limitations section, deployers get an answer on dialect and noise instead of testing for it.
- An independent evaluation that rescores Thonburian Whisper and competing Thai ASR models on one test set with Deepcut would make cross-model claims possible.