Medical Speech-to-Text Software: Compare Transcript Features and Test Accuracy

Medical speech to text software converts spoken clinical language into text that a person can review. The useful output is not simply a paragraph appearing on screen. It is a transcript that preserves what was said, who said it, and enough evidence to correct uncertainty before the text reaches a medical record.

Start with the output you need. A clinician dictating into a text field, a practice reviewing uploaded recordings, and a developer building a transcription application are buying different workflows. A medical model can recognize terminology without providing patient matching, an editing interface, or a reliable EHR save.

This guide focuses on the transcript layer: raw words, speaker labels, settled recognition results, corrections and delivery. For cursor commands and Mac deployment, see the medical voice recognition guide. For preparing the spoken source, see the physician dictation guide. Neither subject needs to be repeated here as another general software ranking.

For a broader platform shortlist, see our medical transcription software guide. For the full capture-to-note workflow, see our AI medical transcription guide. This page concentrates on the raw transcript and API evidence within those workflows.

Sources checked September 30, 2026. The text-scoring exercise was checked on that date; no product audio benchmark was conducted. Examples are fictional, not patient records or supplier outputs. Separate physician review has not been completed. Vero publishes this guide and offers a documentation workspace; vendor documentation below establishes described capabilities, not independent accuracy or superiority. Local clinical, privacy and recordkeeping requirements still govern use.

Choose the transcript route before the brand

There are three useful starting points. Direct text entry fits a clinician who already knows the wording and wants to dictate it into an editor. Recording-to-transcript review fits a workflow in which someone must revisit the source, resolve a difficult word or attribute a statement. A transcription API fits an organization prepared to build and maintain the application around recognition.

A generated note is a different artifact. It may reorganize a conversation, remove repetition, and place information under a template heading. Those transformations can be useful, but they are not verbatim recognition. A well-written note can conceal a bad transcript, and a faithful transcript can contain an incomplete clinician recap.

Write an output contract before a demonstration: “Return a reviewable U.S.-English transcript with speaker roles, recoverable uncertainty and a version identifier,” or “Insert the clinician’s approved dictation into this receiving field.” Specify whether filler words, self-corrections and formatting commands should remain visible or be interpreted. Keep raw recognition evidence separate from the formatted result when the chosen product makes both available.

If the product exposes only a generated note, mark transcript accuracy not assessable through that interface. Do not reverse-engineer a WER claim from the note. Instead, evaluate the note against the source facts, and record the limitation when comparing products. The AI medical dictation guide develops that rewriting distinction.

Documented options: workspaces versus transcription APIs

These four examples illustrate different purchase routes; they are not a tested leaderboard. Findings are vendor-documented, checked September 30, 2026. “Not established” means the reviewed page did not establish the capability—not that the product necessarily lacks it.

Swipe or scroll horizontally to compare every column. Keyboard users can focus the table and use the arrow keys.

Vero fits when the intended endpoint is a reviewed clinical draft with encounter context and a template. Its documentation describes selecting Record, choosing a template, creating the note, reviewing it, then copying or exporting it. That is a useful workflow mechanism, not evidence that its recognition is more accurate than another engine. A requirement for raw, time-aligned transcript export needs a separate verified answer from the supplier. Vero’s note workflow.

Heidi’s documented cursor route and an API’s transcript route solve different problems. For an API, ask who owns the review screen, authentication, failed-upload recovery, storage lifecycle and EHR adapter. For an app, ask which of those functions are included and which remain the clinic’s responsibility. A demonstration should use the actual intended destination, not merely a vendor’s blank editor.

Compare the same transcript features across medical APIs

This is a vendor-documented feature comparison, not locally tested recognition performance, checked September 30, 2026. The Google column is limited to V1’s two medical models; general Speech-to-Text features do not automatically apply to them. AWS batch settings and streaming settings use different names.

Swipe or scroll horizontally to compare both APIs. Focus the table and use the arrow keys for keyboard scrolling.

Speaker partitioning is not channel identification. Partitioning estimates which voice spoke in a mixture; channel identification follows separately captured audio tracks. Neither establishes patient identity or clinical role. In a synthetic telehealth check, map each channel or speaker label to the supplied role, then verify every participant’s utterance is present. A correct label attached to missing audio cannot recover the missing statement.

The practical difference is the review evidence your application must retain. A plain-text export can discard timing, confidence and speaker/channel fields even when the engine returned them. Inspect the raw response and the review screen separately; do not classify an absent UI field as absent engine support, or a configured feature as proven accurate.

What accuracy means at the transcript layer

Use several measurements rather than turning one number into “clinical accuracy.” WER measures lexical differences: substitutions plus deletions plus insertions, divided by the reference word count. Preserve the untouched reference and output, specify tokenization, and state whether case, punctuation, abbreviations and numeric rendering are normalized. Report raw-output and post-correction results separately.

Medical-term errors answer a narrower question. Before running the test, identify each occurrence of the selected specialty terms in the reference. Count altered or missing occurrences against that fixed denominator, and list invented terms separately. Do not select only the terms that survived transcription. A phrase such as “ankle pain” can be one predefined term occurrence while containing two WER tokens; document the distinction.

Assertion errors include incorrect negation, side, source, number, timing and completed-versus-pending status. “Dyspnea” can be spelled correctly while “denies dyspnea” becomes “reports dyspnea.” Attribution errors require a role or channel comparison: matching words do not establish the speaker. Neither check is supplied by a word-only edit distance.

In a 2018 study of 217 clinical notes from two organizations, Zhou and colleagues reported error rates of 7.4% in speech-recognition output, 0.4% after transcriptionist review and 0.3% in physician-signed versions. It is historical evidence that review stage matters, not a contemporary benchmark for the offerings above. The design does not justify assuming that this clinic will achieve those rates or that signing eliminates errors. Original dictated-document study.

Settled text is not verified text

Streaming recognition can expose interim segments that later change. Google distinguishes result stability—whether an interim result may change—from confidence—an estimate of recognition correctness. A final segment is settled recognition output, not a clinical approval. Applications must assemble the complete final transcript rather than export whatever happened to be visible midway through capture. Google’s recognition-result documentation.

For a fictional integration test, display interim text “The report is reviewed,” then replace it with final text “The report is not reviewed.” The destination must use the final result and preserve the negative. A high-confidence label or a stable screen does not excuse inspecting the source assertion. Record which stream event caused export so a premature save can be distinguished from a recognition error.

Worked example: one source, five output states

The fictional source says: “Dysphagia is documented. The patient denies dyspnea. Left ankle pain persists. The outside report is pending.” In the supplied role map, the dyspnea statement is the patient’s report; the remaining wording is the clinician’s recap. These are documentation facts for a text exercise, not a diagnosis or treatment recommendation.

The deliberately flawed candidate changes four words: dysphagia to dysphasia, denies to reports, left to right, and pending to reviewed. Under the displayed scorer’s rules, the reference has 16 words and the candidate has four substitutions: 25% WER. One of the three predefined terms—dysphagia, dyspnea and ankle pain—is wrong, yet all four changes need review.

A transcript score is only one part of the record

Fictional text exercise · no audio or product run

Select a failure stage. Word edits are calculated here; term and role findings are the supplied teaching annotations, not an automated clinical assessment.

Example stateRaw transcriptCorrected transcriptSpeaker label lostOld version savedMissing transcript

Supplied reference · 16 words

Dysphagia is documented. The patient denies dyspnea. Left ankle pain persists. The outside report is pending.

Raw transcript · Raw v1

Dysphasia is documented. The patient reports dyspnea. Right ankle pain persists. The outside report is reviewed.

Word error rate

25.00%

Selected term errors

1/3

Attribution errors

0

4 substitutions · 0 deletions · 0 insertions · 16 reference words

Four word substitutions alter the term, negation, side and report status. Only one of the three selected medical terms is wrong; the unchanged word dyspnea still has the wrong assertion attached to it.

Next action: Replay the source, correct all four differences and preserve the original output for the evaluation log.

See the source-to-text review

• Term: dysphagia must not become dysphasia.
• Negation: denies dyspnea must not become reports dyspnea.
• Laterality: left must not become right.
• Action state: a pending outside report has not been reviewed.
• Speaker map: the dyspnea statement is the patient’s report; the remaining statements are the supplied clinician recap.
• Selected terms: dysphagia, dyspnea and ankle pain. Matching these does not validate their attached assertions.

Copy example evaluation record

Vero text scorer 1.0: case-insensitive word comparison; punctuation and signs are excluded. Speaker and version metadata are checked separately. No text-entry field or audio upload is provided.

The speaker-label state contains exactly the correct words but attaches the patient statement to the clinician. It therefore has 0% WER and one supplied attribution error. The old-version state represents a different problem: someone corrected v1 to v2 in the editor, but v1 reached the receiving record. Reopening the saved copy detects a failure that reviewing v2 alone cannot reveal.

The calculator runs a fixed text comparison, not speech recognition. Its normalizer lowercases text and excludes punctuation and signs. This matters: “−2” and “2” can receive the same word tokens even though their numeric meanings differ. Examine the literal numbers, units and symbols separately. A missing candidate is scored as all reference words deleted; an actual failed session must also remain in the operational results.

A repeatable 32-attempt audio test pack

Use the following as Vero’s proposed editorial test design, not a validated clinical assessment instrument or a completed trial. No microphone, product version, noise recording or recognition output has been tested for this article. An authorized team can run it using synthetic speech, retain the artifacts and publish the observed results with their actual test date.

Run four fixed scripts with two capture configurations, two noise conditions and two repetitions: 4 × 2 × 2 × 2 = 32 attempted recordings per recognition configuration. For example, configuration A can be a built-in laptop microphone at a measured distance; B can be a named USB headset. Use one quiet condition and one fixed, non-patient background-noise file. Preserve that file, playback level and placement. If sound pressure is not measured, say so; “noisy clinic” is not reproducible evidence.

Use the same speakers, scripts and source recordings across the compared engines when upload is supported. For live-only routes, record the changed input conditions and avoid claiming an identical-audio comparison. The repeated fixtures are correlated observations, not 32 independent patients or proof of performance across accents. Add separately reported speakers and languages if those are adoption requirements.

Swipe or scroll horizontally to read each script and expected check.

Freeze both verbatim and formatted references before looking at outputs. For ST-03, recognizing “zero point five” as “0.5” can be semantically correct while differing lexically. Publish which comparison was scored; do not quietly normalize away the defect you intend to detect. For ST-04, a literal transcript and a repaired final sentence can both satisfy different output contracts, but must not be scored against interchangeable targets.

Record date, engine and version if exposed, language/model, specialty, vocabulary list, OS/app/browser, microphone model, input selection, distance, channel count, encoding, sample rate and network conditions. AWS recommends lossless FLAC or WAV with PCM 16-bit encoding and supports sampling at 16 kHz or above for its medical service; do not apply that requirement indiscriminately to every app. AWS input guidance.

Two reviewers should verify the reference against the recording before scoring. A clinician or qualified medical-transcription reviewer should resolve terminology and clinical-assertion disagreements; a technical reviewer should check channels, segment assembly and delivery. Log disputed spans rather than forcing apparent agreement. Report WER using total edits divided by total reference words, plus per-case results and the worst failures. Report term, attribution, numeric and transfer defects separately.

For each attempt, retain the raw result, correction ledger, corrected version, receiving copy and outcome. Time active review/correction and delivery verification; report processing wait separately. Count permission denials, capture loss, empty output, timeout and failed save against all attempted sessions. A low WER among surviving sessions cannot answer whether the application is dependable.

A correction workflow that preserves the evidence

First, isolate the source. Confirm the synthetic or approved encounter, the intended microphone and whether one mixed channel or separate speaker channels are used. A microphone indicator does not establish that every telehealth participant is audible. In a two-speaker rehearsal, have each person say a distinct identifier, then verify both in the transcript before continuing.

Second, preserve raw text and uncertainty. Keep the original output unchanged for evaluation, and create a corrected version. If a word cannot be established from authorized evidence, use the organization’s uncertainty or clarification process. Do not silently choose a plausible medication or condition from context. Speaker diarization can separate voices without proving which voice belongs to the patient; map roles explicitly.

Third, log the correction. A useful row records the source span, raw wording, replacement, error category, reviewer and version. In ST-01, “reports → denies” is a negation repair, not merely style editing. “Pending → reviewed” is a status repair. If a vocabulary change is made, retest a new utterance plus a similar word that must remain unchanged; an exact repeated script alone can conceal overfitting.

Fourth, review downstream transformations. A note generator or formatter may reinterpret the corrected transcript. Google’s V1 medical dictation documentation, for example, describes bracketed spoken-punctuation and formatting-command output. A returned command marker is not necessarily a literal phrase intended for the chart. Define how the consuming application handles it, retain its raw representation and test the visible result. Documented medical-model formatting.

Finally, reopen the receiving record. Compare the saved version with the approved text, including speaker/source wording, numeric symbols and unfinished work. Do not treat clipboard copy, an API success response, or a visible editor as proof of a successful final save. If the wrong version arrived, repair the record under the applicable amendment process and investigate the transfer path. Recognition tuning cannot fix a stale-version save.

Privacy: follow every copy of the recording

Map the path from microphone or file through recognition, temporary storage, transcript history, review screen, generated note, exports and backups. Ask separately about audio and text retention, support access, subprocessors, processing regions, model-improvement use and deletion. A “delete recording” control may leave a transcript, clipboard copy or exported document behind.

For U.S. HIPAA-regulated workflows, HHS explains that a cloud service creating, receiving, maintaining or transmitting ePHI on behalf of a regulated organization is generally a business associate—even when the provider cannot decrypt it. Appropriate agreements and safeguards are required; HHS does not certify individual products as compliant. A supplier badge is therefore not a substitute for reviewing the actual service and configuration. HHS cloud guidance.

Configuration can change the data path. Google says Cloud Speech-to-Text does not log customer audio/transcripts by default, but its optional logging program allows data use to improve the service. The documentation also warns that deleting a project does not delete that logged dataset; a separate deletion request is needed. Verify the project’s settings and applicable terms rather than assuming that account deletion covers every copy. Google’s data-logging documentation.

Patient-recording rules also depend on jurisdiction and setting. Ontario’s CPSO guidance says physicians must inform patients about AI use and obtain consent before recording conversations using AI; it also describes ongoing privacy and professional responsibilities. That is Ontario guidance, not a universal consent form. Have the local organization approve its explanation, refusal/stop workflow and retention plan before recording real encounters. CPSO guidance.

For synthetic evaluation, avoid patient identifiers, real chart exports and background conversations. Assign nonidentifying test IDs and remove the fixtures under the approved schedule. Synthetic content reduces patient-data exposure; it does not waive supplier terms, access restrictions or the need for permission to conduct a competitive product test.

Buying: compare completed work, not audio minutes alone

The best medical speech to text software for a particular clinic is the route it can operate, review and maintain—not necessarily the lowest nominal recognition price. Compare an app subscription per clinician or seat with an API’s billed audio units only after adding the application work around the API.

As of September 30, 2026, Google’s published V1 medical dictation/conversation pricing lists the first 60 minutes per month per account at no charge, then a standard USD $0.078 per minute; commitment discounts are displayed separately. This is not the standard-model rate or a promise of permanently free clinical transcription. The page also says channels are billed separately and storage or other cloud services can add charges. Recheck the exact model, API version and eligible billing terms before purchase. Official pricing.

Here is a hypothetical monthly worksheet, not a supplier quote or measured clinic result. Suppose the billable processing quantity is 1,000 minutes after any applicable allowance, at $0.078/minute: $78. Add $120 for application/support costs, $50 for monthly setup allocation, and 200 transcripts requiring two minutes of review plus one minute of transfer verification at a chosen $60/hour labor valuation: $600. Total: $848, or $4.24 per completed transcript, assuming all 200 were completed.

The $120, $50 and labor inputs are invented planning assumptions. The processing price is a dated published example. If only 180 transcripts are completed after failures while those costs remain unchanged, the total is about $4.71 per completed transcript, before extra recovery effort. For multichannel recognition, substitute the actual billed channel-minutes rather than the recording’s wall-clock duration.

An app comparison should use the same volume, currency, review standard and receiving-record checks. Include seat minimums, unused capacity, microphones, onboarding, support, renewal/cancellation terms and exit export. Labor valuation is not automatically cash savings; saved minutes have operational value only when the clinic can use them. Broad product pricing belongs in the linked dictation software buying guide, while this worksheet tests the transcript route’s true workload.

A practical acceptance decision

Set the decision criteria before testing. Require an authorized data path, adequate coverage of the intended languages and speakers, accessible correction evidence, a dependable receiving-record route, and a staffed fallback. Choose thresholds locally with appropriate reviewers; no universal WER cutoff in this article defines clinical safety.

Then make a narrow decision: accepted for the named synthetic-tested configuration, pilot with a specific unresolved condition, or not accepted. Keep serious identity, permission, missing-audio and false-completion failures distinct from average word performance. A term score should not compensate for the wrong patient or an unavailable transcript.

Publish actual configuration and test dates when the proposed audio pack is run. Until then, the honest finding is documented capabilities plus a checked fictional text exercise, not comparative recognition results. The useful next step is to collect raw output, corrections and the reopened destination for one bounded workflow—not to add an unsupported “most accurate” label.