Skip to main content

Version 2026.11.0

· 12 min read

General changes​

Features​

  • Two health dashboards in Grafana (https://<IP_address_of_virtual_appliance>/grafana) answer "is this appliance healthy?" without having to read Kubernetes
    • Appliance health covers the machine itself: processor and memory use, load average, swap, free space and free inodes on both disks, disk and network throughput, and — on GPU appliances — how many graphics cards there are, their model, memory use, utilisation and temperature
    • Speech Platform health covers the platform running on it: which services are up, anything restarting repeatedly, work in progress and whether it is queueing, spare capacity per technology, and how much disk the platform's working files take
  • Searchable logs in Grafana. The Logs dashboard shows every service on one timeline and narrows by service or by free text, so a problem can be investigated by searching rather than by downloading an archive of /data/log and reading it by hand. Fifteen days of history are kept
  • Serving your own TLS certificate no longer requires editing Kubernetes manifests. Put cert.crt and cert.key into /data/certificate — directly or through File Browser — and the appliance picks them up automatically, with no restart. See Use HTTPS certificate
  • The administration endpoints — File Browser, Grafana, Prometheus and the administration landing page — can be switched off, for deployments that want the Speech Platform reachable while these are not. See Disable administration endpoints

Fixes​

  • Accessing the Virtual Appliance over HTTPS works again. Appliances upgrading from 2026.05.1 and older stopped answering on port 443 after 2026.07.0 replaced the appliance's edge component. The appliance now generates its own self-signed certificate on first boot and terminates TLS with it

Technology changes​

Features​

  • Enhanced Speech to Text Built on Whisper can run several models on one appliance. Upload every model you are licensed for together with its licence and run configure-speech-platform.sh -a. Each model becomes its own instance with its own licence, and a request picks between them with the model parameter; the first instance answers when none is given
    • Note: each model runs as its own instance, so memory use grows with every model you add. Running both large_v2 and large_v3 roughly doubles the memory Whisper needs — RAM on a processor-only appliance, graphics card memory on a GPU one. Size the virtual machine accordingly
  • Voiceprint Extraction and Voiceprint Comparison are merged into the Speaker Identification technology — one technology with one model and one licence instead of two
  • Grapheme to Phoneme Conversion and Phoneme Recognition are now available in every language you have a Keyword Spotting model for, instead of English and Czech only. Neither needs a licence of its own, so the languages you already own become available as soon as the appliance is upgraded

Fixes​

  • configure-speech-platform.sh now configures Deepfake Detection correctly. Autoconfiguration pointed the technology at a licence key its secret does not contain, so the technology came up unlicensed. The technology itself was never affected — it works as before wherever the configuration names the right key

Removed​

  • Referential Deepfake Detection is no longer available. It was meant as an addition to standard Deepfake Detection, potentially offering better accuracy by adding reference authentic audio or a voiceprint, but the standard Deepfake Detection model has evolved to a quality that makes it obsolete

REST API changes​

Features​

  • BREAKING CHANGE: Speech Platform Info lists speaker_identification instead of voiceprint_extraction and voiceprint_comparison. This is the REST API side of the Speaker Identification merge
  • BREAKING CHANGE: A missing language gives a new error. Speech to Text by Phonexia, Keyword Spotting, Phoneme Recognition, Grapheme to Phoneme Conversion and Whisper Speech Translation (output_language) now return 404 with the new detail type language.not-available, whose context has requested_language and available_languages. Before, they returned model.not-deployed with model names and a misleading error detail location. Languages are shown as full codes where the models support them, e.g., cs-CZ instead of cs for Speech to Text by Phonexia. For Speech Translation, en is always listed as a supported output language. The change also affects:
    • An invalid language code now gets a 422 before the availability check
    • Speech Translation now checks that Whisper itself is deployed (not just the required Text Translation model)
    • A non-English output language for Speech Translation built on Whisper with no Text Translation models used to give a 500 and now gives this 404
  • Enhanced Speech to Text Built on Whisper and Speech Translation built on Whisper accept a new model query parameter, which selects between the Whisper models configured on the appliance. The default instance answers when the parameter is omitted (default models are marked with "default": true in the Speech Platform Info response)
  • Every request that starts a technology task accepts an optional task context: any JSON value of up to 64 KiB, such as a call ID, a customer reference or tags, returned in the context field of the Get Task Result response once the task is finished
  • Speech Platform Info now has an optional include_model_features parameter. When set to true, the response will contain additional information about models for technologies that currently support listing model features:
    • Enhanced Speech to Text built on Whisper:
      • languages (language codes for the language parameter in Whisper Speech to Text),
      • translation_languages (language codes for the source_language parameter in Whisper Speech Translation)
    • Speech to Text by Phonexia:
      • graphemes (symbols for specifying spelling for preferred phrases and additional words),
      • phonemes (symbols for specifying pronunciation for additional words),
      • word_separator (symbol for connecting two words in a single spelling for an additional word, e.g., + in English "hello+world")
  • Every Enhanced Speech to Text Built on Whisper result segment now contains the detected_language field. It's the language actually detected, which can differ from language when the language was forced, or when the detected language isn't licensed and the closest licensed one was used
  • GET /api/system/experimental/licenses billing properties have a name when a license splits capacity by operation (transcription or translation for Speech to Text); otherwise the name field is absent. The billing capacity and consumption items for Speech to Text now also carry the operation label (transcribe/translate), so Transcription and Translation are now billed separately

Fixes​

  • Voiceprint Conversion handles voiceprints from audio with no recognized speech. Such a voiceprint used to cause an internal error; now its entry in vector_voiceprints is null
  • Speech to Text by Phonexia rejects a task, instead of failing it, when it reports an input error the REST API doesn't recognize; such errors have the unknown detail type. A word separator at the start or end of an additional word, or inside a preferred phrase, is now rejected with grapheme.invalid
  • Numbers from the technologies have at most 10 decimal places, which removes binary floating-point noise, e.g. 0.30000000000000004 becomes 0.3. This applies to Voiceprint Comparison scores, log-likelihood ratios, Audio Quality Estimation values and Time Analysis speech rate
  • Times are cut to 2 decimal places. This is truncation, not rounding, so an end_time can't go past the end of the recording. Also applies to input times such as the range_from/range_to and max_speech_length query parameters
  • Disabled technologies count as not deployed. A technology that doesn't have any enabled models in the configuration is now treated as not deployed, so a request fails right away with an internal error "Technology models are not deployed." Such technologies are not listed in the Speech Platform Info
  • BREAKING CHANGE: The API health success status changed from 204 to the documented 200 OK
  • Rejected tasks with an unreadable input file now report the single message "Invalid data in input file" instead of "Empty or corrupted media file", "No audio streams found in the file" or "Audio length limit exceeded"
  • language_switching together with source_language no longer fails. Speech Translation used to reject this combination with 422 language-switching-and-source-language-conflict. Now language_switching is ignored when source_language is set (the same way as when language is set for Whisper Speech to Text), and that error detail type is gone

Removed​

  • The experimental Referential Deepfake Detection endpoints (POST /api/technology/experimental/referential-deepfake-detection and GET /api/technology/experimental/referential-deepfake-detection/{task_id}) are gone along with the component, together with the multichannel.not-supported task error detail type, which only these endpoints returned

Deprecations​

  • The X-Location header is marked deprecated in OpenAPI and will be removed in version 2027.1.0
  • GET /api/info/task/{task_id} will also be removed in version 2027.1.0

Documentation​

  • The REST API introduction now says more clearly that new enumeration values may be added to API responses as backwards-compatible changes. Clients should handle values they don't recognize, e.g. by treating them as a generic case
  • The pending and cancelled task states are now documented
  • Speech to Text documentation now says correctly that unsupported graphemes or phonemes lead to a rejected task, not a failed one

GUI (web application) changes​

Features​

  • Added Settings → Diagnostic data section, where you can easily create and directly download a package with diagnostic information needed to investigate and report issues
  • Settings → System Configuration section now allows configuration of multiple models for a technology. This is useful especially for Enhanced Speech to Text Built on Whisper and Speech Translation, which can be now fully managed from the GUI, without editing configuration files.
  • Speech to Text and Speech Translation let you choose the Whisper model. Language and model are offered as one combined option, so each language is listed once for every available model — for example Auto-detect, Built on Whisper (large_v2) and Auto-detect, Built on Whisper (large_v3). Only the models configured on your appliance are shown, and the default model is selected automatically
  • The language, transcription technology and model of recordings already added can be changed from Input parameters in the header of the results pane, using Edit (pencil icon). Changing any of them restarts the transcription for the selected recordings

Fixes​

  • The upload dialog no longer shows an odd file size limit. It read 104.86 MB, because the appliance's limit was set in units the dialog then converted. The limit is now exactly 100 MB and the dialog says so. A limit you have changed on your own appliance is kept as you set it
  • Working with a large number of audio files is now reliable. Uploading and browsing large batches of recordings no longer slows the application down or causes it to become unresponsive
  • Fixed a memory leak triggered when mounting a large amount of audio data. The application's memory use no longer grows with the volume of audio loaded, and memory is released as expected
  • Various smaller improvements across the interface make everyday work in the Speech Platform GUI smoother and more responsive. For example: inputs, selects and dialogs were refreshed for a more consistent look; a dialog whose content could not be scrolled was fixed; a show/hide toggle was added to password field, unified paddings and much more.
Included Components
  • Age Estimation 2.2.0
  • Audio Manipulation Detection 2.2.0
  • Audio Quality Estimation 2.2.0
  • Deepfake Detection 4.4.0
  • Denoiser 2.2.0
  • Emotion Recognition 2.2.0
  • Enhanced Speech to Text Built on Whisper 2.6.0
  • Gender Identification 2.2.0
  • Grapheme to Phoneme Conversion 1.2.0
  • Keyword Spotting 2.3.0
  • Language Identification 2.2.0
  • Media Processing 2.1.0
  • Phoneme Recognition 1.2.1
  • Replay Attack Detection 2.3.0
  • Speaker Diarization 2.2.0
  • Speaker Identification 1.1.0
  • Speech to Text 2.3.0
  • Text Translation 2.2.0
  • Time Analysis 2.2.0
  • Voice Activity Detection 2.3.0