AshvinLohana/FF_blindspot_task

Whisper × Sindhi — A Blind-Spot Probe

A small, self-recorded evaluation set probing a particular failure of open-weight speech models, on low-resource / code-switched Pakistani speech, Whisper large-v3 does not abstain. It confidently misidentifies the language and also fabricates an apparent fluent transcript, and on silence it hallucinates the text.

Created for the Fatima Fellowship "Blind Spots of Frontier Models" challenge by Ashvin Kumar. All audio is my own voice; — the code only logs numbers; I read the Sindhi / Urdu / Gurmukhi transcripts and judged each one by myself.

Answer 1 — about the blind spot challenge

So just last week, our instructor assigned us a research task in class. I asked my colleague to look into it in our native SIndhi language, however he expressed disapproval over its failure to assist in this regard. Later, I investigated it, and found that the model does not know that the audio is in Sindhi or it does not know this language. Once when I asked it in it, it answered about vehicle transportation in a foreign language, and the next time it responded in Urdu. This showed how in the same language and tone, it was still misled about the language and also importantly the message. So basically it confidently fabricated and also appeared to lack absentation to an extent. Benchmarks tend to miss it because typically they score on WER or accuracy on the languages it is trained on and not about level of uncertainty. Although it is true that research has been conducted on how models can be better refined to say “i do not know” such as the R-Tuning method, however the issue is much beyond that. It is quite evident about the model being biased towards popular languages. For instance, the records attached related to my experiment show how confidently it can judge an English (En 8/8) or an Urdu (Ur 8/8) transcript, the latter is my national language, but on Sindhi (Sd 0/11) it easily tends to mix with Punjabi or even Urdu at times. It has also mixed with Gurmukhi script, related to Punjabi. And it is already well known that it fabricates dialogues in silent audios such as saying “thank you” even if the audio may lack words, especially “thank you”. I believe it is a very important concern, although I understand Sindhi is not one of the dominant languages in the world at the moment, but it can really have negative consequences in the long run in a variety of ways, either through false transcripts or miscommunication.

Answer 3 — about a path forward i planned

There can be multiple ways to address this gap. Although in general data curation can help as we have seen in the past AI advancements, here sindhi native speakers like me can work around filtering and assisting with the quality of data, not just quantity of data. Other than this, here we can also follow other strategies such as calibrating abstention. This can be quite useful for the current issue. In the output files attached, some signals can be noticed like no_speech is detected as 0.92 on silence and although low but there is language-ID confidence on Sindhi also. And it tells us that the model appears to ignore these. So we can gate the output on those signals, i.e. abstain or flag the data when confidence is low. This is how we can improve in calibrating abstention and will eventually help us reduce the gap significantly because if it learns to do in this way, it will better process the user inputs and respond more suitably. Furthermore, it is also true that the model is also very confident on the incorrect predictions of language or resulting output so 100% of the time flagging the data may not always help. An example of this from the result set is when G_sd, a sindhi audio, was detected as belonging to Urdu language at 0.96 score, and another audio namely H_sd was recorded to be at 0.94. Here, we can work on another strategy that can help lower this confidence in false prediction. It can possibly be about restructuring the model by giving it incorrect samples possibly and labelling and training it appropriately so it at least discourages itself from highly alien inputs.

---

What's in here

• clips/ — 35 audio clips (.m4a), all my own voice
• sindhi_whisper_eval.csv / .jsonl — per-clip results + my hand labels
• whisper_sindhi_full_eval.ipynb — the G-Colab notebook that produced them

Clip design

• 8 matched trios (A–H): the same content in English, Urdu, and Sindhi — isolates language while holding meaning constant.
• 3 ultra-short Sindhi (short_sd_*) — does short audio break worse?
• 5 code-switch (cs_*) — Sindhi + English, how I actually talk.
• 3 silence / noise (sil_*) — the hallucination probe.

How it was produced

• Model: openai/whisper-large-v3 (open-weight, ~1.5B params), temperature=0.
• Per clip I log: detected language + Whisper's own language-ID confidence (detect_language), the transcript, and two internal signals no_speech_prob and avg_logprob. Language-ID is auto-scored against the true language — Whisper has a Sindhi code (sd), so detecting ur/pa for Sindhi is a genuine miss, not a missing option. Fabrication severity is hand-labelled by me.

Results at a glance

Condition	Clips	Correct language ID
English (control)	8	8 / 8
Urdu (control)	8	8 / 8
Sindhi	11	0 / 11
Code-switch	5	— (no single correct answer; hand-judged)
Silence	3	hallucinated text on 2 / 3

Sindhi was absorbed into Urdu and Punjabi (often written in Gurmukhi / Indian script) — never identified as itself. On silence, no_speech_prob reached 0.92 while the model still emitted "Thank you." Full per-clip detail (confidence, transcripts, my notes) is in the CSV.

Column guide (CSV / JSONL)

• actual — true language I spoke · detected — Whisper's guess · lang_ok — auto match
• det_conf — Whisper's confidence in its own (mis)guess · top3 — its top-3 languages
• no_speech_prob, avg_logprob — Whisper's internal signals
• transcript — what it wrote · i_said — ground truth (mine) · fab_note — my hand judgment

Limitations

Small and deliberately personal: 35 clips, one speaker (me), single recording setup. This is a probe, not a benchmark — meant to expose a failure shape vividly and reproducibly, not to measure population-level error rates.

Reproduce

Open whisper_sindhi_full_eval.ipynb in Colab (Runtime → Change runtime type → T4 GPU), upload the clips/, run top to bottom. It links the model rather than re-hosting it.

References (prior work I build on)

• Zhang et al., R-Tuning: Instructing Large Language Models to Say "I Don't Know" (NAACL 2024), arXiv:2311.09677 — refusal-aware tuning for text LLMs; I ask what the speech analogue looks like.
• Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision (OpenAI Whisper, 2022) — the model under test.

Author

Ashvin Kumar github.com/betheashvin