Is audio-based training or acoustic adaptation available for Norwegian in Azure Custom Speech?

Question

Is audio-based training or acoustic adaptation available for Norwegian in Azure Custom Speech?

Ignacio Massari 0

We are currently working on a real-time speech transcription and translation application using Azure Speech, and we are investigating whether it is possible to improve speech recognition accuracy for Norwegian by training or adapting a Custom Speech model using audio data.

From what we understand, Norwegian (nb-NO) currently supports some Custom Speech capabilities, but audio-based training does not appear to be available in the same way as it is for certain other languages. We would like to clarify whether this is a current limitation or whether there is an existing approach that we may have overlooked.

Our use case is particularly focused on Norwegian dialects and pronunciation differences. Norway has a large number of regional dialects, and the pronunciation of the same words can vary significantly depending on the speaker and region. Because of this, text-based customization alone may not be sufficient for improving recognition of certain dialects or pronunciation patterns.

We would therefore like to know:

Is there currently any way to train or adapt an Azure Speech / Custom Speech model for Norwegian using labeled audio and transcripts?
If audio-based training is not currently supported for Norwegian, is this something Microsoft is planning to support in the future?
Are there any recommended alternatives for improving recognition accuracy across different Norwegian dialects and pronunciation variations?
Is there another Azure Speech feature or model adaptation approach that could help us achieve acoustic or dialect-specific adaptation?

Our main goal is not only to improve recognition of specific vocabulary, but to improve recognition of different pronunciations and dialectal variations of Norwegian speech using real-world audio examples.

Any clarification or recommendations from the Azure Speech team would be greatly appreciated.

Thank you.

0 comments

2 answers

Your answer

Answer 1

For Norwegian (nb-NO), audio-based custom speech training is not currently supported.

Supported custom speech capability for nb-NO is:

Plain text
Output format

Audio + human-labeled transcript training is not listed for Norwegian, while it is listed for many other locales. That means there is currently no supported way to train or adapt an Azure Custom Speech model for Norwegian by using labeled audio and transcripts for acoustic adaptation.

On future support, no roadmap or announcement is provided in the available documentation.

Recommended alternatives from the available Azure Speech guidance are:

Use plain-text custom speech customization for Norwegian to improve recognition of domain-specific vocabulary.
Use phrase list where supported as a runtime recognition feature to improve recognition accuracy for anticipated words and phrases. Phrase list does not require model training.
If overall speech-to-text accuracy remains low for the scenario, custom speech is the documented model customization approach, but for Norwegian that customization is limited to the capabilities listed for nb-NO.

For the specific goal of improving recognition of dialect and pronunciation variation, the available documentation states that audio + human-labeled transcripts are the dataset type used to improve acoustic aspects such as slight accents, speaking styles, and background noises. Since Norwegian does not list audio-based customization support, that acoustic adaptation path is not available for nb-NO at this time.

Answer 2

**Welcome to Microsoft Q&A!**Hello @Ignacio Massari I hope you are doing well,

For Norwegian Bokmål (nb-NO), the current Azure Speech documentation lists plain-text customization, but not audio + human-labeled transcript training. Therefore, acoustic adaptation using labeled Norwegian audio is not currently supported. ([Microsoft Learn](https://learn.microsoft.com/azure/ai-services/speech-service/language-support?tabs=pronunciation-assessment&utm_source=chatgpt.com"Language and Voice Support for Azure Speech - Foundry Tools | Microsoft Learn"))

For your dialect-focused scenario, the practical options are:

Use plain-text Custom Speech to improve domain-specific vocabulary.
Use phrase lists at runtime to boost names, locations, terminology, and other expected phrases. ([Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/improve-accuracy-phrase-list?utm_source=chatgpt.com"Improve recognition accuracy with phrase list - Foundry Tools | Microsoft Learn"))
Test representative recordings from the different Norwegian dialects to measure where the base model struggles.
Microsoft's documentation explains that audio + human-labeled transcripts are specifically useful for improving acoustic characteristics such as accents and speaking styles, but that capability depends on the locale's supported customization features. nb-NO currently isn't listed as supporting that training method. ([Microsoft Learn](https://learn.microsoft.com/en-sg/azure/ai-services/speech-service/language-support?tabs=tts&utm_source=chatgpt.com"Language and Voice Support for Azure Speech - Foundry Tools | Microsoft Learn"))

References:

If this answer helped clarify the platform capabilities and save you troubleshooting time, please consider marking it as Accepted. This helps others in the community find similar solutions.