Follow new updates and improvements to Speechmatics.
Oak 1, a new medical model, is now available for batch transcription. It delivers improved medical keyword recognition compared to Melia 1, with the same base accuracy.
A new Melia 1 model improves transcription accuracy across many languages, including a 1.6% word error rate improvement in English and stronger recognition of medical terminology.
Faster transcription for Melia 1.
Updated Orchestrator Version: 2026.10.14+8bef15f3f2+1.11.0
Version 16.2.0 is now available for Batch Container, Realtime Container, GPU Transcription Inference Container and GPU Translation Inference Container.
Improved accuracy for GPU Enhanced English when transcribing recent years (e.g. "2026").
Fix for English transcription with the Enhanced model: spoken numbers like "ten thousand" were sometimes incorrectly transcribed as "ten zero zero zero".
Improved accuracy and latency for Dutch with the Enhanced model when processing non-speech audio.
Fixed slight accuracy degradation with ForceEndOfUtterance for below languages, by reverting a previous latency optimisation, for: ba, be, bg, bn, ca, cmn, cmn_en, cs, cy, de, el, eo, et, eu, fa, gl, he, hi, hr, hu, ia, id, it, ko, lt, lv, mn, mr, ms, pl, pt, ro, sk, sl, sw, ta, th, tr, ug, uk, ur, vi, yue. ForceEndOfUtterance still works for all languages.
Software Bill of Materials (SBOM) is available for download from the corresponding release page in our Support Portal.
OpenVEX statements are also available for each image, download from the corresponding release page in our Support Portal. These encode our assessment of CVEs that don't affect this release.
Triton Inference Server updated to 26.07, addressing the previously-tracked CVE-2026-47627 (Critical), CVE-2026-47628 and CVE-2026-47629 (High) against the GPU Transcription Inference and GPU Translation Inference images.
How to read these tables. Each row is keyed by the package a scanner attributes the advisory to. That attribution comes from the distribution's own security tracker rather than from the advisory's subject, so where an advisory concerns a component bundled by a different project, the tracker may attribute it to a package this image does ship even though the vulnerable component itself is not present. Scanner coverage also differs: not every row below is reported by every tool.
The following vulnerabilities are fixed in the versions this release ships. Scanners that identify a package by the distribution's own records will continue to report them, because those records track the distribution's package version rather than the version installed here.
The following vulnerabilities were reviewed and determined to have no security impact for this release.
The following vulnerabilities were reviewed and do apply to this release, with no fix released upstream at the time of writing. They are tracked and will be taken up when a fix becomes available.
Melia 1, our multilingual speech-to-text model, is now available in Realtime.
Melia 1 automatically detects the spoken language across 56 languages, so you no longer need to set one up front. It also handles code-switching at the word level, so speakers can change languages mid-stream. Every transcribed word includes a language label as metadata.
Available in US, EU and Australia regions.
Already know the language? Set language_hints configuration to one or more language codes when opening the WebSocket connection to bias recognition toward them.
Adds a new message LanguageInfo which is sent by the server the first time each language is spoken to format the received text, see the API reference for more detail.
Get started: Set model=”melia-1” and language=”multi” with the latest Python or JavaScript SDKs or see API documentation here.
Updated Orchestrator Version: 2026.09.30+1aeef15746+1.9.0