Release Notes

Follow new updates and improvements to Speechmatics.

July 20th, 2026

2026.07.20 Realtime SaaS

New

A new global endpoint is now available at global.rt.speechmatics.com. This address automatically routes connections to the nearest region, providing the lowest latency and eliminating the need to hard-code specific regional endpoints into your applications.

Regional endpoints such as eu.rt.speechmatics.com and us.rt.speechmatics.com remain available for customers with specific data residency requirements.

Refer to the Realtime transcription documentation for full details.

July 15th, 2026

Realtime SaaS

2026.07.15 Realtime SaaS

New

Realtime transcript messages (AddTranscript , AddPartialTranscript, and EndOfUtterance) now include a forced flag, allowing clients to identify transcripts generated by a forced end of utterance. See the documentation for more details.

Improvements

Improved accuracy (reduced deletions at end of utterance) when using ForceEndOfUtterance, with specific values of max_delay.

Updated Orchestrator Version: 2026.07.10+d3875410b2+15.15.0

July 14th, 2026

Batch SaaS

2026.07.14 Batch SaaS

Improvements

New Melia 1 model for multilingual transcription provides:

  • Improved English transcription accuracy (WER reduced by 2.4% relative)

  • Improved code-switching accuracy with language hints

  • Improved alphanumeric formatting

Updated Orchestrator Version: 2026.07.09+8723c03fc9+1.4.0

July 10th, 2026

Realtime Kubernetes

1.4.0 - Realtime Kubernetes

Added

  • Added support for Gateway API HTTPRoute and ListenerSet resources for the proxy service.

    • Gateway API support can be used as an alternative to Ingress and an Ingress controller.

    • Gateway resources are not created by this chart and must be managed separately.

Updates

  • Updated the Redis dependency to use 8.6.4-alpine image.

  • Updated the Realtime container and GPU Transcription Inference Container version to 15.14.0. For more information, see the 15.14.0 container release notes.

  • Updated the sessiongroups CustomResourceDefinitions and controller.

    • Refer to the sm-realtime Helm chart RELEASE_NOTES.md for upgrade details.

Deprecated

  • Deprecated the proxy.ingress.url Helm value in favour of proxy.ingress.hostname for configuring the Ingress hostname.

    • Existing deployments remain compatible: if proxy.ingress.hostname is not set, the chart will continue to use proxy.ingress.url as a fallback.

Security fixes

Vulnerability Management

  • 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.

Non-Applicable CVEs

The following vulnerabilities were reviewed and determined to have no security impact on this release due to specific configurations or the use of closed, trusted environments.

Container

Component

Identified CVEs

Reason

Transcriber (Realtime)

libgstreamer

CVE-2025-3887

Manually patched (build 1.24.2-sm-1320260); scanners may still flag it because the base version string remains unchanged

Transcriber (Realtime)

ffmpeg and bundled codec libraries (libopenexr, libmfx1, libde265, libzvbi, libopenh264, libopenjp2, libvo-amrwbenc, libsndfile, libsoup, libcurl, libp11-kit, glib-networking)

Multiple CVEs - see attached OpenVEX file for full per-CVE list

False positives: most stem from scanners matching against an older ffmpeg version tracker rather than the newer Speechmatics build (8.1.1) actually shipped; the rest are codec paths (video encode/decode, teletext, JPEG2000, etc.) never invoked by our audio-only decode-to-WAV pipeline

Transcriber (Realtime)

transformers

CVE-2026-4372, CVE-2026-1839

Model-loading / checkpoint-deserialization RCE path not invoked at runtime - the deployed code never calls from_pretrained / AutoModel

Transcriber (Realtime)

python3-pip (vendored urllib3/requests)

CVE-2024-35195, CVE-2025-66418, CVE-2025-66471, CVE-2026-21441, CVE-2026-9375

Build/install-time only, not on any runtime path

GPU Inference Server (Transcription)

wheel, setuptools

CVE-2026-24049, CVE-2024-6345, CVE-2025-47273

Base-image packaging tooling not invoked at inference runtime; fixes available, to be picked up in a future rebuild

The control-plane components and bundled Redis image were also scanned (Medium severity and above) and no High or Critical vulnerabilities were identified.

July 7th, 2026

Batch Container

GPU Transcription Inference Container

1.3.0 - Melia 1 Containers

Version 1.3.0 is now available for the Melia 1 Container (Batch Transcriber and GPU Transcription Inference Container).

New

Introducing Melia 1 - our brand new model for multilingual transcription capable of handling production-scale traffic for multilingual use cases. Refer to the documentation for more details.

Melia 1 allows us to transcribe any of our 55+ languages without having to specify languages up front. Accuracy is already superior to our Standard model while offering lower processing costs.

Key features

  • Multilingual transcription with language labelling

  • Code-switching between languages, particularly when switching between English and Spanish, Arabic, Mandarin, Malay or Tamil.

  • Optional language hints

  • Blazingly fast turnaround time

Security fixes

Vulnerability management

  • 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.

Non-applicable CVEs

The following vulnerabilities were reviewed and determined to have no security impact for this release.

Container

Component

Identified CVEs

Reason

Transcriber

libgstreamer

CVE-2025-3887

Manually patched (build 1.24.2-sm-1320260); scanners may still flag it because the base version string remains unchanged

Transcriber

ffmpeg and bundled codec libraries (libopenexr, libmfx1, libde265, libzvbi, libopenh264, libopenjp2, libvo-amrwbenc, libsndfile)

Multiple CVEs — see attached OpenVEX file for full per-CVE list

False positives: most stem from scanners matching against an older ffmpeg version tracker rather than the newer Speechmatics build (8.1.1) actually shipped; the rest are codec paths (video encode/decode, teletext, JPEG2000, etc.) never invoked by our audio-only decode-to-WAV pipeline

Transcriber

transformers

CVE-2026-4372, CVE-2026-1839

Deployed code only uses tokenization (AutoTokenizer/AutoConfig) on bundled files; the vulnerable model-load path is never invoked

Transcriber

python3-pip (vendored urllib3/requests)

CVE-2024-35195, CVE-2025-66418, CVE-2025-66471, CVE-2026-21441, CVE-2026-1703

Build/install-time only, not on any runtime path

Transcriber

libsoup-3.0-0 / libsoup-3.0-common

CVE-2025-14523, CVE-2025-32049, CVE-2025-4035, CVE-2026-0716, CVE-2026-0719, CVE-2026-3632, CVE-2026-3633, CVE-2026-3634, CVE-2025-9901, CVE-2026-3099

Not used by any pipeline; outbound HTTP uses the Python requests/urllib3 stack

Transcriber

libp11-kit0

CVE-2026-13757

p11-kit RPC server/proxy not used

Transcriber

glib-networking family

CVE-2026-10028

TLS backend reached only via libsoup, which no pipeline uses

Transcriber

libcurl3t64-gnutls

CVE-2026-11856, CVE-2026-10536, CVE-2026-12064

No SMB usage and no gstreamer-curl plugin in any pipeline; outbound HTTP uses the Python requests/urllib3 stack

GPU Inference Server

libcurl3t64-gnutls

CVE-2026-11856, CVE-2026-8925, CVE-2026-8927, CVE-2026-10536, CVE-2026-12064, CVE-2026-8286, CVE-2026-8458, CVE-2026-8924, CVE-2026-9547

Orphan package (empty reverse-dependency set); no installed runtime component links or invokes it

GPU Inference Server

setuptools, wheel

CVE-2024-6345, CVE-2025-47273, CVE-2026-24049

Base-image packaging tooling not invoked at inference runtime

GPU Inference Server

transformers

CVE-2026-4372, CVE-2026-1839

Only the fast tokenizer is used; the checkpoint-deserialization/model-load path is never invoked

July 6th, 2026

Batch Container

Realtime Container

GPU Transcription Inference Container

GPU Translation Inference Container

15.14.0 - Containers

Version 15.14.0 is now available for Batch Container, Realtime Container, GPU Transcription Inference Container and GPU Translation Inference Container.

New

GPU & CPU

  • Added a new model property in transcription config for selecting Standard or Enhanced models. See the documentation for more details.

GPU

  • Added support for NVIDIA Blackwell architecture GPUs.

Improvements

GPU

  • New Enhanced model for French (fr) and Russian (ru) give up to 2% relative accuracy improvement.

  • New Enhanced model for German (de) improves accuracy on numerical digit sequences, formatting consistency of alphanumeric entities, and punctuation and numerical entity interactions, resulting in more consistent formatting.

GPU & CPU

  • New Standard and Enhanced models for Norwegian (no) improve the formatting of alphanumeric entities, with additional support for correctly handling:

    • Traditional compounded cardinals

    • Approximate ranges

    • Blood pressure measurements

  • Improved the handling of ForceEndOfUtterance to avoid suppressing zero-duration final transcripts.

Deprecated

GPU & CPU

  • The operating_point property in transcription config is deprecated in favour of the new model property for model selection. See the documentation for more details.

Security fixes

Vulnerability management

  • 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.

Non-applicable CVEs

The following vulnerabilities were reviewed and determined to have no security impact for this release.

Container

Component

Identified CVEs

Reason

Transcriber (Batch, Realtime)

libgstreamer

CVE-2025-3887

Manually patched (build 1.24.2-sm-1320260); scanners may still flag it because the base version string remains unchanged

Transcriber (Batch, Realtime)

ffmpeg and bundled codec libraries (libopenexr, libmfx1, libde265, libzvbi, libopenh264, libopenjp2, libvo-amrwbenc, libsndfile, libsoup, libcurl, libp11-kit, glib-networking)

Multiple CVEs - Refer OpenVEX file for full per-CVE list

False positives: most stem from scanners matching against an older ffmpeg version tracker rather than the newer Speechmatics build (8.1.1) actually shipped; the rest are codec paths (video encode/decode, teletext, JPEG2000, etc.) never invoked by our audio-only decode-to-WAV pipeline

Transcriber (Batch, Realtime)

transformers

CVE-2026-1839

Checkpoint-deserialization path not invoked at runtime

Transcriber (Batch, Realtime)

python3-pip (vendored urllib3/requests)

CVE-2024-35195, CVE-2025-66418, CVE-2025-66471, CVE-2026-21441, CVE-2026-1703

Build/install-time only, not on any runtime path

GPU Inference Server (Transcription and Translation)

wheel, setuptools

CVE-2026-24049, CVE-2024-6345, CVE-2025-47273

Base-image packaging tooling not invoked at inference runtime

‌

July 2nd, 2026

Realtime SaaS

2026.07.02 Realtime SaaS

New

  • Added a new model property in transcription config for selecting Standard, Enhanced, or Melia-1 models. See the documentation for more details.

Improvements

  • Improved the handling of ForceEndOfUtterance to avoid suppressing zero-duration final transcripts.

Deprecated

  • The operating_point property in transcription config is deprecated in favour of the new model property for model selection. See the documentation for more details.

Updated Orchestrator Version: 2026.06.28+6f4c629d3a+15.14.0

July 1st, 2026

Batch SaaS

2026.07.01 Batch SaaS

New

  • Added a new model property in transcription config for selecting Standard, Enhanced, or Melia-1 models. See the documentation for more details.

Deprecated

  • The operating_point property in transcription config is deprecated in favour of the new model property for model selection. See the documentation for more details.

Updated Orchestrator Version: 2026.06.28+6f4c629d3a+15.14.0

June 30th, 2026

Batch SaaS

2026.06.30 Batch SaaS

New

Added support for Synchronous transcription, enabling transcripts to be returned in a single request without polling. Refer to the documentation for more details.

Improvements

New Melia 1 model for multilingual transcription provides:

  • Improved code-switch accuracy when not using language hints. Particular improvements when switching between English and Spanish, Arabic, Mandarin, Malay or Tamil.

  • Improved formatting of alphanumerics

  • Improved overall transcription accuracy

Updated Orchestrator Version: 2026.06.28+6f4c629d3a+1.3.0

June 24th, 2026

Realtime SaaS

2026.06.24 Realtime SaaS

Improvements

New Standard and Enhanced models for Norwegian (no) improve the formatting of alphanumeric entities, with additional support for correctly handling:

  • Traditional compounded cardinals

  • Approximate ranges

  • Blood pressure measurements

Updated Orchestrator Version: 2026.06.17+2e6888fd18+15.13.0