Japanese Voice-to-Text: Preventing Unwanted Kanji Transcription for Names

Thierry Tropée 40 Reputation points
2025-02-14T06:06:11.7233333+00:00

When using Azure Speech to Text for batch transcription of conversations in Japanese, there is an issue with person names being transcribed into incorrect Kanji characters. A custom speech model has been created to handle specific industry terms, but names continue to be converted into unwanted Kanji, which complicates transcript searches. Is there a way to prevent names (like ドイ) from being transcribed into Kanji characters (such as 土井 or 土居)?

Is it possible to train a speech model to address this issue? If so, what steps or examples should be followed to achieve this?

Azure Speech in Foundry Tools
Answer accepted by question author
Pavankumar Purilla 11,655 Reputation points Moderator
2025-02-15T00:22:56.1866667+00:00

Hi Thierry Tropée,
Greetings & Welcome to Microsoft Q&A forum! Thanks for posting your query!

To prevent unwanted Kanji transcription of names when using Azure Speech-to-Text for Japanese batch transcription, you can leverage a combination of custom pronunciation dictionaries, custom speech models, and post-processing techniques.
Initially creating a custom pronunciation dictionary allows you to specify how names should be transcribed, ensuring that names like "ドイ" remain in Katakana rather than being converted into Kanji such as "土井" or "土居." This can be done in Azure Speech Studio by adding pronunciation rules that explicitly map names to their desired form.

If pronunciation dictionaries alone do not resolve the issue, training a custom speech model using labeled audio data where names are consistently written in Katakana can help. This involves collecting training data, uploading correctly transcribed text, and training a model to reinforce the desired transcription format.
I hope this information helps.

Was this answer helpful?

1 person found this answer helpful.

0 additional answers

Sort by: Most helpful
Most helpful Newest Oldest

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.