A consultation can move between languages without anyone announcing the change. A patient may describe a symptom in one language, name a medicine in English and ask a relative to clarify what happened. A useful medical scribe must preserve the meaning of that exchange, not simply translate its most recognisable words.
EkaCare's Parrotlet-a 2.5 Pro release page is dated 10 August 2026. Its developer documentation describes medical speech recognition for Indian English, Hindi, Marathi, Kannada and Telugu, including code-mixed speech. The company identifies its Parrotlet-a models as part of the EkaScribe documentation stack.
That makes the release relevant to clinicians considering multilingual documentation. It does not make every supported language equally reliable in every setting, or a speech-recognition benchmark equivalent to a validated clinical note.
Distinguish the model from the finished scribe
Automatic speech recognition, often shortened to ASR, converts audio into text. A documentation product must do more: manage the recording, organise the conversation, generate the requested note and provide a usable review workflow.
This distinction changes the buying question. A clinician assessing EkaScribe should examine the final documentation experience. A technical team assessing Parrotlet-a should examine the model's performance and operational requirements. Neither assessment can simply borrow the conclusion of the other.
The Eka developer description presents a stack that moves from medical speech into structured documentation. The appropriate evaluation should follow the same journey. Where did an error enter? Was a word misheard, a speaker confused or a correctly transcribed statement incorrectly summarised?
Without that separation, a team may try to fix the wrong component. Improving speech recognition will not necessarily resolve a note-generation rule that routinely drops uncertainty or merges statements from different people.
Test the consultation that actually happens
Imagine a fictional consultation in which a patient describes intermittent discomfort in Hindi, the clinician asks questions in English and a family member adds an observation in Marathi. The medicine name is spoken quickly, while the patient corrects the timing of a previous episode.
The evaluation should not ask only whether the system recognises each language. It should ask whether it preserves the correction, attributes the relative's statement appropriately and distinguishes what the patient experienced from what the clinician is considering.
A monolingual test clip could miss all of those challenges. So could a neatly scripted dialogue in which speakers wait their turn and never interrupt. A useful local test set should deliberately reflect the pace, recording conditions and conversational structure of the intended service.
This does not require publishing private recordings or uploading patient data to an unapproved demonstration site. Teams can begin with consented, authorised or simulated material under their own governance arrangements. The point is to test realistic structure without treating convenience as permission to disclose information.
A word error is not always a clinical error, and vice versa
Word error rate is a measure of transcription differences. It can be useful, but the clinical consequence depends on what changed. A harmless variation in phrasing and the loss of a negation should not be treated as interchangeable merely because both alter the text.
The Parrotlet-a model card includes developer-reported semantic word error comparisons and audio-quality analyses. These provide questions worth investigating, rather than a reason to assume the model is the best choice for every consultation.
For a clinical evaluation, reviewers could classify errors by their effect: changed meaning, lost uncertainty, incorrect attribution, omitted information or additional unsupported information. They should also record whether the error is easy to spot during ordinary review.
Consider another fictional example: a transcript retains most of a sentence but omits "not". Its overall word agreement may remain high, while the statement's meaning reverses. Conversely, a transcript might use a different but equivalent phrase that does not alter the clinical account. The evaluation must make room for that distinction.
The benchmark needs a boundary around its claims
Before accepting a performance ranking, ask which recordings were used, how the reference transcripts were created and how closely the test resembles your service. Check whether the results are produced by the developer, an independent group or a local evaluation team.
A model may perform differently across recording environments, consultation lengths or language combinations. The appropriate question is not whether the published benchmark is useless. It is what decision that benchmark can reasonably support.
Versioning also matters. A result applies to the evaluated model and workflow, not automatically to every subsequent release. A hospital should retain the tested version, evaluation date and relevant configuration alongside its findings.
For EkaScribe, the more persuasive evidence would connect speech performance with note quality and clinician correction effort. For Parrotlet-a, it would include reproducible evaluation of the model under the intended conditions. These are complementary evidence requirements, not competing definitions of accuracy.
Review the note for clinical relationships
A final note needs to preserve relationships as well as individual terms. Who reported the symptom? Was a condition confirmed or considered? Did the clinician change the plan after hearing new information? What did the patient decline or remain uncertain about?
A proposed local review could compare the note against a structured account of the encounter prepared independently by a clinician. Reviewers would identify both missing statements and changes in how the statements relate to each other.
They should also assess the documentation format. A concise note may be preferable to a verbose one, provided it preserves the important distinctions. More text is not necessarily more faithful, and a beautifully organised output should not receive credit for information the conversation never established.
This is particularly relevant when language switching occurs around a correction or qualification. The sentence that changes the interpretation may not be the longest or most easily recognised part of the recording.
Make the trial useful for the clinician
A practical trial should record how long review takes, which errors recur and whether corrected documents can be filed without further reformatting. It should include cases where the clinician chooses not to use the output and ask why.
For example, a system might be helpful in straightforward consultations but require substantial checking when several people speak. That would support a narrower initial use case rather than an all-or-nothing verdict. Explicit limits are more useful than a broad declaration that a tool "supports" a language.
The comparison should also preserve the distinction between documentation and reference. A scribe records an encounter; a referenced question investigates a clinical uncertainty. Ask-iatroX addresses the latter workflow and should not be portrayed as a multilingual transcription engine or an automatic validator of EkaScribe's output.
The strongest case for EkaCare's language-specific work is that it invites evaluation of the consultation as it actually occurs. The conclusion should come from preserved meaning, usable notes and manageable correction, not a count of languages on a product page.
Frequently asked questions
Is Parrotlet-a 2.5 Pro the same product as EkaScribe?
No. Parrotlet-a is part of the speech-recognition technology described by EkaCare. EkaScribe is the documentation product and associated developer offering built around a broader workflow. The developer overview distinguishes these layers.
Which languages does the current model documentation list?
The model documentation checked on 16 October 2026 lists Indian English, Hindi, Marathi, Kannada and Telugu, including code-mixed speech. Listed support should still be tested in the intended clinical setting.
Does a strong transcription benchmark establish that the final note is clinically accurate?
No. Errors can arise during attribution, selection and summarisation after transcription. The final note and the review workflow need their own evaluation, even when the underlying speech model performs well.
Explore referenced clinical questions alongside your documentation workflow →