OmniScribe 2 is Omniloy's self-hosted transcription system, and it uses context from the patient's record (in this test that context included the medical terms spoken in each clip: a best case). On 300 Spanish clinical clips with synthetic voices it gets 85.7% of medical terms right. The six other systems, three commercial APIs and three open-source models in their default setup, get 50–71%. On clean audio it also has the lowest word error rate of all seven. On degraded audio (noise, phone codecs, reverb), Soniox has a lower word error rate and Cohere Transcribe is level. OmniScribe 2 can also write a look-alike drug name instead of the one spoken.
Each system is compared with OmniScribe 2 using a paired bootstrap on the same clips. ours better and theirs better mean the 95% interval of the difference excludes zero; level means it includes zero. Terms right counts a term only if every word of it is correct.
OmniScribe 2 used context from the patient's record; in this test that context included the medical terms spoken in each clip: a best case. Clean: 147 clips. Degraded: 153 clips with noise at 20 to 5 dB SNR, and on many clips also a phone or low-bitrate codec, room reverb, or a speed or pitch change. Every other system ran in its default configuration: no custom vocabulary, keyterms or context prompt, and Spanish selected wherever the system allows it (Voxtral Realtime has no language setting). Empty outputs count as errors: scribe_v2_medical returned no text on 9 clinical clips, Voxtral Realtime on 7 and nova-3-medical on 1. Confidence intervals resample whole sentences, because the same sentence appears in several clips.
Each line runs from the 147 clean clips to the 153 degraded ones. OmniScribe 2 stays ahead of every system on medical terms in both conditions. On word error rate it is the lowest of the seven on clean audio. On degraded audio Soniox is lower and Cohere Transcribe is level, and that is where Soniox’s overall WER lead comes from.
The clean and degraded groups are mostly different sentences, so each line compares two groups of clips, not the same clips getting worse.
In 38 of the 300 clips, OmniScribe 2 gets every medical term right and all six other systems miss at least one; the reverse happens on 3. Against Soniox alone, OmniScribe 2 gets every term where Soniox misses one on 66 clips, and the reverse happens on 19.
OmniScribe 2 had patient-record context that included the spoken terms, a best case, while the other six ran in their default setup.
On three real-speech sets together (FLEURS, VoxPopuli and MediaSpeech: real speakers, general Spanish, no medical terms), OmniScribe 2 has 6.6% average WER. That is level with Cohere Transcribe (6.8%), which is better on VoxPopuli and worse on MediaSpeech, and significantly ahead of the other five systems on this page (7.4–9.4%), by a paired bootstrap on the three-set average. In the full leaderboard, two open models that are not on this page, omniASR-LLM-7B (6.5%) and Voxtral Small 24B (6.6%), have a slightly lower or the same average.
On degraded clips OmniScribe 2 makes 3 or more errors on 47 clips, Soniox on 37. The cases below are clips where Soniox makes at most 2 errors and OmniScribe 2 at least 4 more, plus one clean clip. The recognizer can drop the rest of a sentence or replace it with fluent Spanish that was never said. The clinically serious failure is different: it can write a look-alike drug name instead of the one spoken. This happened on one clean clip too. The next things to fix are noise-robust input, a guard that catches cut-off or invented output, and tighter handling of look-alike drug names.
An inserted word is one that appears in the transcript but was never said. On the clinical set, OmniScribe 2’s rate of inserted words is level with Soniox, nova-3-medical, Voxtral Realtime and Cohere, and lower than Whisper-turbo. It also inserts fewer words than scribe_v2_medical (10.3 vs 10.6 per 1,000 words), a gap small enough to count as level. On FLEURS real speech it has the lowest rate of the seven (level with Whisper-turbo and Voxtral Realtime). Its distinctive weakness is the one shown above: occasionally writing a look-alike drug name.
Per 1,000 reference words, scored the same way for every system; labels use a paired bootstrap against OmniScribe 2. VoxPopuli and MediaSpeech are left out because their references skip some real speech, for example a Greek phrase at the start of one MediaSpeech clip, and any system that transcribes it is counted as inserting words.
Three widely used open-weights models ran on the same clips: OpenAI’s Whisper large-v3-turbo, Mistral’s Voxtral Mini 4B Realtime and Cohere Transcribe. On clinical audio, OmniScribe 2 gets 85.7% of medical terms right against 60–64% for the three. On real human speech, Cohere Transcribe is the strongest of the three: level with OmniScribe 2 on average, better on VoxPopuli parliament speech and worse on MediaSpeech. The full leaderboard has 22 open models; the best of them on clinical audio, Voxtral Small 24B, has the same clinical WER as OmniScribe 2 (8.5%) and gets 70.3% of medical terms right.
Open models run offline, like OmniScribe 2, in their default configuration: no custom vocabulary, keyterms or context prompt, and Spanish selected for Whisper and Cohere (Voxtral Realtime has no language setting). Whisper accepts a text prompt, which was not part of this comparison. Voxtral Realtime returned empty output on 7 clinical clips.