Introducing Medera BH-ASR, and the leaderboard we hold it to

A therapy or primary-care conversation is a hard case for speech recognition: two people, long pauses, clinical vocabulary, and now and then a sentence about safety that a clinician must not miss. Medera BH-ASR is built for that setting, in English and Arabic. Today its model card and a live benchmark are public on Hugging Face, so anyone can see how it is built, how it ranks against open recognizers, and where it still falls short.

One recording in, a reviewed transcript out

Nothing is declared up front. From a single recording, BH-ASR finds the speech, tells the speakers apart, detects the language of each stretch, and works out who is the clinician and who is the patient. It returns a versioned JSON transcript, bh-transcript.v1, with speaker, role, language, guard flags and safety flags on every segment.

1. 1

   Speech regions

   Voice-activity detection with Silero VAD 6.2.

2. 2

   Speakers

   TitaNet-large voice embeddings from short windows, clustered by cosine distance.

3. 3

   Language

   Detected per speech region with Whisper-large-v3-turbo, so a conversation can move between English and Arabic.

4. 4

   Recognition

   Pro fuses Parakeet-TDT-0.6B-v2, Qwen3-ASR-0.6B and Cohere Transcribe word by word in English, and uses Cohere Transcribe for Arabic. Lite runs one recognizer: Parakeet-TDT-0.6B-v2 for English, Nemotron-3.5-ASR for Arabic.

5. 5

   Verifier and guard

   Whisper-large-v3-turbo transcribes independently. A guard compares the fused text with it and keeps, flags, trims, replaces or drops each segment, recording the reason.

6. 6

   Roles and safety language

   Roles are read from the words themselves. Bilingual rules surface safety language as prompts for the clinician to review.

Everything runs on CPU, and every model file is pinned by SHA-256. BH-ASR trains no recognizer weights of its own: it is a system built around open models, which keep their own licenses. The engine, the fusion, the guard and the rules are ours.

The leaderboard

Thirteen systems ran on the same CPU runtime and were scored on 480 clips across three test sets: 240 clips of real English mock primary-care consultations from PriMock57, and 120 English and 120 Arabic clips of scripted safety statements in synthetic speech. The leaderboard was scored on October 8, 2026.

Read the top of the table with care. Pro's English lead over Qwen3-ASR 0.6B is not statistically significant (p = 0.15), so we read it as at the top, not as a margin. What is significant is the fusion: Pro is 1.3 points below Parakeet-TDT-0.6B-v2, the recognizer it is built around (p = 0.004). Resampling whole consultations rather than clips, Pro's English word error rate is 11.02%, with a 95% confidence interval of 9.16% to 13.25% across 42 consultations.

In Arabic, Pro has the lowest word error rate on the safety statements, 35.35%, which is 1.2 points below Cohere Transcribe running alone (p = 0.036). On real read speech from FLEURS (ar_eg), the lighter profile does better: Lite scores 11.14% against 12.38% for Pro (p = 0.027).

Guarded against invented text

Recognizers sometimes write words nobody said, and in a clinical record an invented medication is worse than a missing one. Given a medication prompt, the prompted recognizer on its own inserted 71 drug names that were never spoken. Pro inserted none. On 68 real non-speech clips, Pro produced text on 1, against 4 for Parakeet-TDT-0.6B-v2 alone and 0 for Nemotron-3.5-ASR.

Safety language, surfaced for a clinician

BH-ASR's bilingual rules look for five categories: suicidal thoughts, suicidal plans or means, self-harm, harm to others, and hopelessness. Affirmed, hypothetical and about-someone-else statements become review prompts. Denials and a clinician's own screening questions are documented, not prompted. None of this is a diagnosis or a risk score: severity is set by category alone and has not been validated as a measure of risk.

We test the rules on 60 scripted statements per language, 22 of them alerts, each recorded twice, clean and with room noise. End to end, through recognition, Pro keeps 12 of 22 English alerts and 4 of 22 Arabic alerts. The useful part is where the rest are lost.

In English, recognition loses nothing: the rules are the limit, and indirect phrasing is most of what they miss. In Arabic, recognition is the limit. That tells us where to work first.

A conversation-reading version of the rules, still in development, reads each answer against the clinician's question. On 40 alerts per language in held-out written exchanges, rules 1.5.0 raised 27 English alerts (6 when read line by line) and 19 Arabic alerts (4 line by line), with one false prompt in 24 safe exchanges per language. It is measured on text with the speakers given, not yet through recognition.

Fast enough on a CPU, and private

A real-time factor below 1.0 is faster than the audio itself. On two virtual CPUs, Lite transcribes English consultations faster than they were spoken. Pro takes about twice as long as the recording, which is why it is meant for the transcript after the session. Audio stays in memory, the engine makes no outbound connection, and no recording is kept. For Pro, plan a node with 12 to 16 GB of memory.

Try it, and check our numbers

The live benchmark

Record or upload a conversation of up to 3 minutes in English or Arabic. BH-ASR detects the language, the speakers and their roles, then every open recognizer on the Space transcribes the same speech and is scored live by how much they agree. Accuracy against human transcripts is on the leaderboard. It is a demonstration, not a clinical tool: please don't upload recordings of real patients.

Open the live benchmarkRead the model cardRequest API access

What it is not, yet

• Not a medical device, and not for diagnosis, screening, risk assessment or triage. It is a Beta for clinician-in-the-loop documentation and review.
• English accuracy is measured on role-played primary-care consultations (PriMock57). No real behavioral-health session audio has been evaluated yet.
• The safety test sets are scripted statements in synthetic speech, so those numbers are relative comparisons. Arabic review prompts are not yet reliable through recognition.
• The rules have not yet been reviewed by bilingual clinicians, and every flag is a prompt for review, never screening.
• Speakers are told apart by voice, which fails when voices are alike, and children's speech is untested.
• Vocal biomarkers are in development and not evaluated. The live benchmark returns none of them.

What comes next

• A context model whose evidence is added to the rules. It can add review prompts, never remove one.
• A miss rate you can count: a threshold set from calibration alerts, which needs at least 19 of them for a 5% target.
• A hard-case program with recordings from consenting speakers, not patients: indirect phrasing, alerts buried in long answers, crying or whispered speech, and Arabic dialects.
• Vocal biomarkers measured against each speaker's own baseline, computed only where a deployment enables them.

If you can help with hard cases, as a speaker or as a clinical reviewer, we would like to hear from you. Get in touch.

A note on the numbers