The Clinical ASR Becomes the Language Gate

Not a Medical Device Claim

The paper, arXiv:2606.26901 [cs.CL, cs.AI], was submitted on June 25, 2026. arXiv lists the exact title as SamaVaani: Auditing and Debiasing Multilingual Clinical ASR for Indian Languages, by Subham Kumar, Prakrithi Shivaprakash, Abhishek Manoharan, Astut Kurariya, Diptadhi Mukherjee, Prabhat Chand, Pratima Murthy, Koustav Rudra, Lekhansh Shukla, and Animesh Mukherjee.

This is not a product clearance, clinical trial, or instruction to use automatic speech recognition in care. It is an audit of transcription performance and fairness in real psychiatric interviews. That distinction matters because the paper's object is not the diagnosis itself. It is the path by which spoken care becomes text, evidence, training material, and downstream medical record.

The Psychiatric Interview Corpus

The authors frame psychiatry as unusually dependent on detailed interviews rather than laboratory or radiological tests. They also note that manual transcription is strenuous, error-prone, and can take 5-8 hours for each hour of audio. In psychiatric settings, a transcription error can change the clinical interpretation, especially when speech contains pauses, hesitation, fast speech, repetition, code-switching, background noise, or disorder-specific language patterns.

The study uses 202 audio recordings from a tertiary teaching hospital, collected in real wards and outpatient rooms rather than in an engineered quiet booth. The corpus contains speech from 130 unique speakers: 7 doctors or therapists and 123 patients. The languages are Indian English, Hindi, and Kannada. The paper states that written informed consent was obtained and that the data cannot be made public even after deidentification because it contains sensitive personal health information.

Eight ASR Systems

The audit compares eight ASR systems: IndicWhisper, WhisperLargeV3, Sarvam, GoogleS2T, Gemma3n, OmniLingual, Vaani, and Gemini. The paper says GoogleS2T, Sarvam, and Gemini are proprietary API-inferenced models, while the others are open-source. Some systems accept long-form audio directly, while several are run on 30-second chunks.

The main performance measure is word error rate, or WER, with substitution, deletion, and insertion components. The fairness analysis looks across language, speaker role, gender, and error patterns. That is already a governance choice: the paper refuses to treat one average transcript score as the whole story.

What Broke

The headline result is not that every model failed. It is that performance moved sharply with language and social position. Gemini had the best reported WER among the audited systems: 14.15% for English, 18.52% for Hindi, and 35.01% for Kannada. Even the best system therefore made Kannada transcription far harder than Indian English transcription in this setting.

The authors report substantial variability across systems and languages, with some models performing competitively in Indian English but failing or struggling in regional speech. Kannada is consistently harder, and models such as WhisperLargeV3, GoogleS2T, and Gemma3n show very high WER for Kannada. The paper also reports systematic performance gaps tied to speaker role and gender. In a clinic, that means the error is not merely technical noise. It can decide whose voice becomes legible to the institution.

What SamaVaani Changes

The authors then fine-tune two of the best performing open-source models, Gemma3n and OmniLingual, using train, development, and test splits stratified by language. The paper's train set contains 83.16 hours of audio, the development set 9.64 hours, and the test set 10.22 hours.

SamaVaani is the proposed fairness-aware method. It combines contrastive learning and CTC alignment to improve transcription while reducing subgroup gaps. The authors report up to about 50% WER reduction and consistent fairness-score improvements of 13-41% compared with standard fine-tuning. They also state that SamaVaani still struggles most with Kannada, like the other models. The useful lesson is not that debiasing solves clinical listening. It is that a model can improve and still require subgroup receipts.

Governance Reading

For AI governance, the clinical transcript is not just an output. The paper describes psychiatric transcripts as material for clinical diagnosis, academic training, qualitative research, and AI-system development. If the ASR layer hears clinician speech more reliably than patient speech, or hears high-resource language better than regional speech, then the record can privilege the easier voice before any human reviewer sees the note.

The control is not a general ban on ASR. It is a demand for traceable listening. A deployed system should preserve the original audio access policy, model name and version, language identification, diarization method, WER tests by language and role, subgroup error analysis, human correction path, and rules for downstream reuse. This belongs beside AI audit trails, language access layers, and the existing evidence that speech systems can listen unevenly.

Limits

The paper's limits are important. Its evaluation covers Indian English, Hindi, and Kannada psychiatric interviews, not all Indian languages, dialects, clinical specialties, or code-mixed settings. The authors also say their LoRA setup used rank 8 because of hardware constraints and that full supervised fine-tuning might further improve performance.

WER is not the same as clinical meaning. A small word error can be harmless, or it can matter enormously if it changes medication, time, negation, symptom, speaker, or intent. The privacy limit is also central: deidentified clinical audio can still be too sensitive for public release.

Transcript Receipt

A clinical-ASR receipt should record: language, setting, microphone path, model version, chunking or long-form handling, diarization method, WER tests, subgroup gaps, uncertain spans, human correction authority, consent scope, retention period, and downstream use. The audit-grade sentence is simple: this transcript is a machine-mediated clinical record whose listening errors are measurable and correctable. The weaker sentence is the familiar one: the model transcribed the visit.

Sources

• Subham Kumar, Prakrithi Shivaprakash, Abhishek Manoharan, Astut Kurariya, Diptadhi Mukherjee, Prabhat Chand, Pratima Murthy, Koustav Rudra, Lekhansh Shukla, and Animesh Mukherjee, SamaVaani: Auditing and Debiasing Multilingual Clinical ASR for Indian Languages, arXiv:2606.26901 [cs.CL, cs.AI], submitted June 25, 2026.
• arXiv PDF and HTML versions: PDF and experimental HTML, reviewed for title, authorship, date, subject categories, dataset size, language coverage, ASR model list, WER results, SamaVaani method, fairness-score claims, consent/privacy statements, and stated limits.
• Related pages: The AI Scribe Becomes the Medical Record, The Machine Interpreter Becomes the Language Gate, The Accent Filter Becomes the Labor Mask, The Health LLM Becomes the Black-Box Evaluation, AI Governance, and AI Audit Trails.

Return to Blog