BatchRecognition returns English text for Japanese audio with no documented language setting

[hrtk91]

Describe the bug

In my diagnostic test, Microsoft.Windows.AI.Speech.BatchRecognition returned English text that appears to be a translation of Japanese speech instead of a Japanese transcription.

The English output broadly follows the meaning of the Japanese input, although some details are mistranslated or omitted.

I could not find a language or transcription-vs-translation option on SpeechRecognitionModel, BatchRecognition, or RecognizeFromFile. The Windows AI Speech documentation and the Microsoft AI Dev Gallery sample use the same API shape without a language setting.

For one reproducible input, the spoken Japanese is:

北浜鉄道は、若葉線の新型車両を来月十五日から運行します。車内には車椅子用のスペースと、百ボルトの電源コンセントを設置しました。

The result returned by RecognizeFromFile is:

Kitahama Railway transports the new model of Wakabasen from the 15th of next month. Inside the car, we installed a space for the car and a 100 volt power supply.

The WAV was generated with the Microsoft Haruka Desktop Japanese voice. A native Japanese speaker listened to the WAV and found the speech clearly intelligible and natural, without anything that sounded like non-Japanese speech. It is passed to RecognizeFromFile as mono, 16 kHz, 16-bit PCM. I can provide the WAV if it would help with investigation.

Steps to reproduce the bug

1. Create a packaged MSIX application with the systemAIModels capability.
2. Prepare the speech recognition model with SpeechRecognitionModel.EnsureReadyAsync() if required.
3. Use a mono, 16 kHz, 16-bit PCM WAV containing the Japanese sentence above.
4. Run the following relevant code:

var modelResult = await SpeechRecognitionModel.TryCreateAsync();
if (modelResult.SpeechModel is null)
{
    throw new InvalidOperationException(
        $"Failed to create SpeechRecognitionModel: {modelResult.ExtendedError}");
}

using var recognition = new BatchRecognition(modelResult.SpeechModel);
string text = await recognition.RecognizeFromFile(wavPath);

I also tested one run with both of the following applied before creating the model:

• declaring ja-jp in the MSIX resources;
• requesting ja-JP through Windows.Globalization.ApplicationLanguages.PrimaryLanguageOverride (the value read back inside the packaged process was normalized to ja).

The control and language-override runs returned exactly the same 161-character English output.

In a small single-voice diagnostic set of 12 synthetic Japanese-news WAV files, 9 results were English-dominant. All files used the same Microsoft Haruka Desktop voice at several speaking rates. The Windows culture, UI culture, and system locale were all ja-JP.

Expected behavior

For Japanese speech, I expected BatchRecognition to return a Japanese transcription rather than English text that appears to be a translation.

If automatic language detection and translation are intentional, I would expect a documented way to select Japanese and request transcription instead of translation.

Could you clarify whether there is currently a supported way to constrain recognition to Japanese and disable translation?

Screenshots

No response

NuGet package version

2.2.2-experimental9

Packaging type

Packaged (MSIX)

Windows version

Windows 11 version 25H2, OS build 26200.8655 (General Availability Channel; not enrolled in Windows Insider)

IDE

Other (dotnet CLI, .NET 9)

Additional context

• Microsoft.WindowsAppSDK.AI: 2.2.6-experimental
• Installed speech component: Microsoft.Windows.AI.Speech.Preview.CPU 1.2607.112.0 (x64)
• OS architecture: x64
• The input is transcoded to the same canonical format used by the Microsoft AI Dev Gallery Speech Recognition sample: mono, 16 kHz, 16-bit PCM WAV.