🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦
Dissecting Qwen3-ASR-0.6B: One Neuron Pushing the Model Toward Thai Script, and the Bug-vs-Gap Lesson
Quay lại Blog

Dissecting Qwen3-ASR-0.6B: One Neuron Pushing the Model Toward Thai Script, and the Bug-vs-Gap Lesson

18 tháng 11, 20268 phút đọcVũ Khôi
English
#English
Vũ Khôi

Vũ Khôi

Tác giả tại Wotech Electronics Blog

Neuron map visualization: neuron 128 at layer 22 of Qwen3-ASR-0.6B with a weight column pushing the model toward Thai, Chinese, and Cantonese

Dissecting Qwen3-ASR-0.6B: One Neuron Pushing the Model Toward Thai Script, and the Bug-vs-Gap Lesson

The audio was "xin chào gia đình Tanaka", and the model returned a Thai-script sentence that reads almost like the Vietnamese sounds. Author Nam Nam, in the Machine Learning Cơ Bản group, "dissected" the open-source speech recognition model Qwen3-ASR-0.6B to find the cause — and traced it to a suspicious neuron at layer 22. But the most readable part of the story is not the neuron; it is how he distinguished a bug from a gap before fixing anything, and how careful he was with his conclusion.

Quick answer: The Qwen3-ASR-0.6B model heard the Vietnamese sounds fairly well but picked the wrong alphabet, returning Thai script for "xin chào gia đình Tanaka". The author determined this was a bug, not a gap, because at 12 of 18 checked positions the correct token was already in the top 10 at the middle layers — the answer was already in the model's head. After scanning 1,050 neurons at layers 18–24, the suspect emerged: neuron 128 at layer 22, whose weight column pushes the model toward Thai, Chinese, and Cantonese. Hooking that neuron to suppress it only at language-deciding tokens fixed 8/8 error sentences, broke none of 8 clean sentences, and cut WER on the trial set by nearly 4% of errors. But writing directly into the weights failed: deleting the neuron's column fixed 0/8, and the bias approach broke 20/207 correct sentences. The author stresses the causality check covered only 8 error sentences and this error type is only about 2% of sentences, so he does not call neuron 128 the model's "language switch".

A language tag is what Qwen3-ASR generates before transcribing: the model decides the language first, then turns audio into text in that language. If the tag is wrong, the whole sentence follows it — which is why an error at the language-choosing step can turn an entire Vietnamese sentence into Thai script even though the model heard the sounds correctly.

The model heard the sounds right, so why Thai script?

The author tells it like this: the audio was "xin chào gia đình Tanaka", and the model returned ชินแจว จาริน ตาณกา. At first he thought it was garbage output, but reading it in Thai sounds gave almost "chin-chao cha-rin ta-na-ka" — the model heard quite a lot of the sounds, but picked the wrong alphabet. Forcing it to Vietnamese gave "Chinh chào, Charing Tanaka": not fully correct, but much closer to the original sentence.

Qwen3-ASR-0.6B is an open-source speech recognition model from the Qwen team, about 0.8 billion parameters, supporting 30 languages including Vietnamese, small enough to run on a personal machine — which is why the author likes tinkering with it. This Thai-script error, as he describes it, is not the model being "deaf" but the model "hearing clearly yet writing in the wrong script".

Is this a bug or a gap?

Before fixing anything, the author asked himself a question that decides the treatment: is this a bug or a gap? By his definition, a gap means the model has not learned it yet — that needs more data and training. A bug means the model already knows, but something is blocking it — that can be fixed directly inside.

Most of the model's Vietnamese errors, he measured, are gaps. But the Thai-script error is different: at 12 of 18 positions he checked, the correct token was already in the top 10 at the middle layers. In other words, the answer was already in the model's head; it just did not win at the final selection step. That evidence led him to classify the Thai-script error as a bug — and a bug can be intervened on directly inside the model.

How was neuron 128 at layer 22 found?

Qwen3-ASR has a habit: it generates the language tag first, then transcribes. If the tag is wrong, the whole sentence follows. The author put an LLM to "examine" the model together with a logit lens, attribution, and a neuron-explorer-style neuron scan that has become popular recently. After scanning 1,050 neurons at layers 18 to 24, the suspect emerged: neuron 128 at layer 22. Its weight column pushes the model toward Thai, Chinese, and Cantonese.

In other words, his method was systematic scanning rather than guessing: measuring which neuron, at which layer, influences the language decision — and only then intervening at exactly that spot, instead of blindly editing the whole model.

How effective was the hook — and what about editing the weights?

The author tried a hook to suppress neuron 128, only at language-deciding tokens. The full comparison is as follows:

Intervention Result How the author reads it
Hook suppressing neuron 128 at language-deciding tokens Fixed 8/8 error sentences, broke none of 8 clean sentences; WER on the trial set down nearly 4% of errors Intervening at the right time and place works without breaking anything
Random direction in the same layer (control) Fixed only 1.6/8 on average, and broke 2.2/8 Shows the hook's effect is not from random intervention — only the right neuron works
Writing into weights: deleting the neuron's column Fixed 0/8 Deleting it outright has no effect
Writing into weights: bias approach Fixed the errors, but broke 20/207 correct sentences, and WER went up Fixes the errors but breaks more than it fixes

The lesson the author draws: a hook can choose when to intervene, while weights act at every step, including when the model is transcribing normally. The same neuron, intervened on selectively by timing, fixes the error; intervened on permanently in the weights, it breaks more than it fixes.

How much should this result be trusted?

The author sets his own limits on the conclusion, and that is the most notable part of the post. He states clearly: causality was checked on only 8 error sentences, the hook results have not yet been run on a test set, and this error type accounts for only about 2% of sentences in the set he measured. So he does not dare call neuron 128 the model's "language switch" — he can only say it plays a real role in this error.

That caution is what separates a serious experiment from a sensational claim: one neuron playing a real role in one specific error, on a small check set, is not enough to conclude it is the switch controlling the model's entire language-choosing behavior. The post leaves open further verification on a larger test set.

Frequently asked questions

1. What is Qwen3-ASR-0.6B? According to the post, it is an open-source speech recognition model from the Qwen team, about 0.8 billion parameters, supporting 30 languages including Vietnamese, small enough to run on a personal machine.

2. How does the Thai-script error happen? The model hears the audio "xin chào gia đình Tanaka" and returns Thai script that reads almost like the Vietnamese sounds. According to the author, the model hears quite a lot of the sounds but picks the wrong alphabet, because it generates the language tag first and then transcribes — a wrong tag drags the whole sentence with it.

3. How do bug and gap differ in the author's framing? A gap means the model has not learned it yet and needs more data and training. A bug means the model already knows the answer but something blocks it, so it can be fixed directly inside. The evidence that the Thai-script error is a bug: at 12 of 18 checked positions, the correct token was already in the top 10 at the middle layers.

4. How effective was the hook suppressing neuron 128? According to the post, the hook fixed 8/8 error sentences, broke none of 8 clean sentences, and cut WER on the trial set by nearly 4% of errors. A random-direction control in the same layer fixed only 1.6/8 on average and broke 2.2/8.

5. Why not edit the neuron's weights directly? Because weights act at every step, including when the model transcribes normally. According to the post, deleting the neuron's column fixed 0/8, while the bias approach fixed the errors but broke 20/207 correct sentences and raised WER. The hook only intervenes at the language-deciding moment.

Source: post by Nam Nam in the Machine Learning Cơ Bản group (18/11/2026). The experiments, figures, and interpretations in this article belong to the post's author and have not been independently verified by Wotech.