Next-generation speech recognition models and infrastructure
New and improved model architecture and infrastructure bring the following actions and benefits:| Actions | Benefits |
|---|---|
| Optimize STT performance | Maximize accuracy of transcripts and diarization while minimizing hallucination risk |
| Refine system configuration | Improve (reduce) latency, augment formatting, and re-introduce interim results feature |
| Infrastructure resilience | Improve system scalability and GPU resource utilization |
/transcribe and /stream endpoints:| Language | Language Code |
|---|---|
| Danish | da |
| English | en, en-GB |
| French | fr |
| German | de |
| Swiss German | de-CH |
Interim Results now available during real-time dictation for low latency transcript previews
Use theinterimResults configuration parameter in /transcribe real-time dictation requests to have transcript previews returned from the server at a faster rate than final transcripts.See full feature details here, parameter configuration details here, and availability by language here.Locale-based formatting for dates, times, and numbers
Dictation formatting can now apply proper styling based on local standards for dates, times, and numbers. Locale is determined based on theprimaryLanguage defined in the web socket configuration.See full formatting details per language here and parameter configuration details here.Note that legacy date parameter values are still supported by the API; no breaking changes were implemented.
Dictation formatting improvements
- Expand units handling
- Expand
- Improve vertebrae formatting: Support for both
letter numberandletter number letter numberpatterns (e.g., “L three” -> “L3” and “L three L four” -> “L3-L4”) - Improve TNM and cancer stage formatting: Support for
T number N number M numberandstage numberpatterns (e.g., “T two N one M one a” -> “T2 N1 M1a” and “stage two b” -> “Stage IIB”) - Improve percentage handling: Move handling from
spokenPunctuationtoformattingso that percent symbol is only returned when there is a number measurement (e.g., dictation, “What percent sure are you question mark eighty percent” -> “What percent sure are you? 80%”) - Expand handling of German dates: Years without spaces (e.g. “zweitausendsechsundswanzig” -> “2026”) and new pattern (e.g., include pattern “siebter fünfter neunzehn vierundachtzig” -> “07/05/1984”)
- Expand handling of German units: Units with -n ending (e.g., “Millilitern”)
- Add new default for ordinals,
numerals_above_nine: Ordinals one through nine are written out (first, second, third) and ten and above are abbreviated (10th, 11th, 12th)
Dictation formatting improvements
- Improved support for
en-GBregional spelling variations - Improved support for
de-CH,gsw-CHregional spelling variations - Remove extra whitespace observed with degree and percent symbols
- Updated handling of single digit numbers, now represented as numeral when followed by year/ month/ week/ hour/ day/ minute/ second(s)
- Updated handling of hyphenated numbers and age dictation
- Updated handling of “one twenty” (and similar number pattern) dictations
- Updated handling of Month-Year dictation
- Fix for an issue that occasionally caused both formatted and unformatted versions of transcript text to be emitted
Update to handling of number formatting in Danish.
Bug fixes:
- Update to French formatting for improved number handling
- Update to Danish formatting for liter abbreviation handling
- Bug fix for extra whitespace being returned with commands or punctuation
- Command service performance and reliability improvements
Improved handling of initialisms in German dictation
An updated version of the German (de) language model was released with improved handling of dictated initialisms: spoken letters, numbers, and abbreviations. Over 800 additional medical abbreviations were introduced to the system, with focused improvement on dental tooth exams, adding support for ICD-10 codes, and ability to have individual letters/numbers returned from the model.Audio input validation
Improved server-side validation of streamed audio - if audio does not meet requirements outlined here, then an Invalid Audio400 error will be returned.Addition of new web socket event: flush
Provide clients the ability to force clear the audio buffer without closing the connection. In response to a "type": "flush" message from the client, the server will return recognized text and/or commands and respond with "type": "flushed", and keep the web socket connection open.See more details on /transcribe here and /stream here.Update to Swiss German language codes
There are now two different language codes that may be used for Swiss German use cases:- Swiss German (language code
gsw-CH), where dialectical Swiss German is spoken (recommended for AI scribe use cases) - Swiss High German (language code
de-CH), where Swiss High German is spoken (recommended for dictation use cases)
Updated Danish language model
Updated Danish (da) language model is now available for dictation (/transcribe) and ambient (/stream) use cases. The language is rated as premier tier as over 158,000 medical terms were included in the training and validation data sets.Bug fixes:
- Fixed an issue that prevented
automaticPunctuationparameter from working as expected. Furthermore,spokenPunctuationandautomaticPunctuationare mutually exclusive: Only one of these parameters should be set totruein a given/transcribeconfiguration, and if both settings are present and set to true, thenspokenPunctuationwill take precedence. See more detail here.
Updated French language model
Updated French (fr) language model is now available for dictation (/transcribe) and ambient (/stream) use cases. The language is rated as premier since over 170,000 medical terms were included in the training and validation data sets.Dictation now supported in Dutch and Norwegian
Norwegian (no) and Dutch (nl) are now available for use on the /transcribe endpoint for dictation use cases. As a result they have been moved to the enhanced language tier. Additionally, Hungarian (hu) medical terminology has been expanded and improved.See more information about supported functionality for dictation here and language tiers here.Feature limitation:
- Some of the recent
/transcriptsendpoint language model updates (de,en,en-GB,fr) have been rolled back so that performance of asynchronous audio file processing can be improved. Infrastructure updates are underway to improve model performance. - Speech-to-text accuracy from
/transcriptsasynchronous audio file processing may be degraded as compared to real-time audio processing via the/streamand/transcribeAPIs, which are not impacted by this issue.
Updated English language model
Updated English (en) language model is now available for dictation (/transcribe), ambient (/stream), and transcription (/transcripts) use cases. This version is rated as premier tier as over 170,000 medical terms were included in the training and validation data sets.Improved transcoding support in speech-to-text APIs
Update to both streaming and asynchronous audio file processing so that transcoding is supported. Previously audio files were required to conform to 16-bit, 16kHz formatting. Now, assuming proper file types are used, any precision and sample rate are accepted. See more details here.Parameter deprecation:
- Definition of the parameter
modelNameis no longer required in/transcriptsAPI requests. The latest and greatest model available per language will be applied automatically. - If the argument is included in the request, then it will be ignored. If the configuration is otherwise valid, then the request will process as expected. See full specification here.
Feature limitation:
- The interim (preview) results feature of the
/transcribeAPI is not performing as expected. As a result, it is being disabled for most languages to prevent issues with speech-to-text latency and accuracy. - If the
interimResultsparameter is included in the request for a language that does not support this functionality it will be ignored and, so long as the request is otherwise valid, the configuration will be accepted.
Conversational transcripts now supported in Arabic
Arabic (language codear) is now available for ambient documentation workflow - Capture conversations spoken in Arabic via the /stream API.New dictation functionality available: Formatting
New dictation functionality available: Formatting. Take control over how dates, time, units, and numbers should be transcribed in the streaming speech to text output.See detailed API specification here and documentation of all available formatting options here.Updated German language model
Updated German (de) language model is now available for dictation (/transcribe), ambient (/stream), and transcription (/transcripts) use cases. This version is rated premier tier as over 150,000 medical terms were included in the training and validation data sets.Dictation now supported in Hungarian
Hungarian (hu) Enhanced language model is now available for dictation (/transcribe).Swedish (sv) Enhanced language model is now available for dictation (
/transcribe).Updated Norwegian (no) and Swedish (sv) Base language models are now available for ambient documentation (
/stream) and transcription (/transcripts) workflows.Danish (da), German (de), and French (fr) Enhanced language models now available for dictation (
/transcribe) workflows.2025-04-09
Swiss German (de-CH) Enhanced language model now available for ambient documentation (
/stream) workflows.French (fr) Enhanced language model now available for ambient documentation (
/stream) workflows.Introducing a new tier system for defining functionality and performance of speech to text language models. Read more about it here.
Updated German (de) and Swiss German (de-CH) language models available.
Updated Danish (da) and Swedish (sv) language models available.
Announcing the launch of Corti AI Platform. Read more about it here.The following languages are supported by Corti speech to text: English (en for US English, and en-GB for UK English), Danish (da), German (de), Swiss German (de-CH), French (fr), Swedish (sv), Spanish (es), Norwegian (nl), Dutch (no), Italian (it), and Portuguese (pt)