Audio + transcript failing to export from Editor to Training and Testing Dataset in Azure Speech Studio

Amy Highton 0

I'm having a really hard time uploading audio and transcripts in Speech Studio. I tried many times to upload audio + human-labelled transcripts and kept getting these errors:

"Your Custom Speech data failed to upload. Check the supported data formats and try again later." AND "Zero transcriptions could be parsed from the given input. Error: invalid input line format..."

I then tried to use the Editor tab instead and upload only the audio, and have Azure automatically transcribe it. After editing the results I tried to export it back to the Training and Testing Dataset tab and got this message:

"Error: normalized text is empty."

I'm sure I have met all of the file requirements, but nothing seems to be working - any ideas?

santoshkc 8,860 Reputation points Microsoft Vendor

2023-10-30T09:16:06.0733333+00:00

Hi @Amy Highton ,

Thank you for using Microsoft Q&A forum!

Can you please provide more details about the file formats, the steps you are taking to upload the files and which Azure documentation on Speech Studio you been following?

Thank you!
Amy Highton 0 Reputation points

2023-10-30T20:12:56.04+00:00

Sure thing: I've been uploading a 16hz mono WAV file of <1min audio along with a .txt file saved as UTF-8 BOM. I've been using this documentation although I wasn't completely sure about formatting - does each new sentence need to be on a new line, and does each line need to have the audio file name even if it's from the same audio piece? When this failed a few times I tried using the Editor tab, following this - but I haven't had any luck with exporting it back to the main training tab - this is where I started getting the "normalised text is empty" message. Let me know what other info I can provide!

1 answer

santoshkc 8,860 Reputation points Microsoft Vendor

2023-10-31T12:02:32.44+00:00
Hi @Amy Highton ,

Thank you for your response.

I tried to reproduce the error that you received.

To fix this kind of issues, see the below steps for best practice:

Use good quality of audio and follow the below given information of image: The uploaded file should be compressed zip file. In the zip file include audio(.wav) and transcript(.txt) file.

In text transcript(.txt) file, it should be audio name + space(or tabspace) + text transcript. eg: audio1.wav is the audio name. see below image.

Upload your data in Training and testing dataset tab. Editor tab is used only to import the data.

Hope this helps. Thank you!

If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.
Please sign in to rate this answer.
Amy Highton 0 Reputation points

2023-10-31T20:23:18.27+00:00

I eventually had success in Editor once I realised that the way I was editing the automated transcription was the problem - I was introducing punctuation that the system didn't accept. But when I tried to replicate this in the 'Training and testing dataset' tab (literally I pasted in the exact transcript and audio file that uploaded successfully in Editor) I got this error: "Error: audio file does not exist". I haven't changed anything about the audio file - any idea what could be causing this?

santoshkc 8,860 Reputation points Microsoft Vendor

2023-11-01T12:21:10.5833333+00:00

Hi @Amy Highton,

Thank you for your response.

Though if you are introducing punctuation. In the text file please include the "audiofile name along with .wav extension (e.g. audio1.wav) followed by tabspace or space and audio text".

Still if you are facing any issue. I request you to raise a support case through Azure portal.

I hope you understand. Thank you.

https://portal.azure.com/#view/Microsoft_Azure_Support/HelpAndSupportBlade/~/overview

YutongTie-MSFT 51,731 Reputation points

2023-11-15T19:01:00.0233333+00:00

Hi @Amy Highton,
Please let us know how above comment works for you. Thanks a lot.

Regards,

Yutong

Amy Highton 0 Reputation points

2023-11-16T20:42:03.52+00:00

Hi Yutong, I'm no longer having any issues using the 'Training and testing dataset' tab - in fact, I seem to be getting away with leaving almost all of the original punctuation in the text and it still uploads fine! I also heard from another team that it was possible to use audio files longer than 1min so I've started using 5min files - do you know if there might be any consequences to this down the line? I'm not seeing any errors reported but in the documentation it says those files will be skipped over? The only issue I'm having at the moment is one error report I can't figure out as there is no file name information to go on - any ideas? I can't see anything problematic in the text

filename: inputText: ... message: Error: invalid input line format: status: Failed

santoshkc 8,860 Reputation points Microsoft Vendor

2023-11-17T11:22:21.6133333+00:00

Hi @Amy Highton

I'm glad to hear that you're no longer having issues with the Training and Testing Dataset tab and that you're able to upload audio files. Regarding the consequences of using longer audio files, it's important to note that longer audio files may take longer to process and may require more resources. However, as long as you're not experiencing any errors or issues, it should be fine to use longer audio files.

Regarding the error report you're seeing, it seems like there is an issue with the input line format. It's possible that there is some formatting issue with the text that is causing the error. You can also try to upload a different file to see if the issue persists.
If you continue to experience issues, you can raise a new query regarding your issue.

I hope you understand! Thank you.

Amy Highton 0 Reputation points

2023-11-19T19:18:05.5033333+00:00

Thanks for the reply - oddly enough, at some point during the uploading process, the error disappeared! Mysterious. Good to have that reassurance re audio files though. There's actually another issue I'm having: I'm uploading a lot of Māori language text for my model and I noticed that it is randomly capitalising a lot of the smaller words, e.g. "tena koe E te mangai O te whare". I tried to export it to the editor and fix it up, but when I returned it to the testing tab the capitalisation was back, and in some cases worse! Is there any way to stop this happening or fix it up permanently when it does? We were told in our initial setup meeting with Microsoft that Māori is a supported language, although we were always unsure about this as it's not in the language list for speech-to-text specifically. Thanks!

santoshkc 8,860 Reputation points Microsoft Vendor

2023-11-20T11:29:01.41+00:00

Hi @Amy Highton,

It's great to hear that the previous issue has been resolved. Regarding the capitalization issue, I don't have any knowledge of Māori language and also this language is not in the language list.
Speech Studio from capitalizing certain words, you can try adding the words to a custom dictionary. The custom dictionary allows you to specify the pronunciation and capitalization of words that aren't recognized by the Speech Studio's default language model. You can add the words to the custom dictionary by following these steps:

Go to the "Custom Speech" tab in the Speech Studio.

Click on the "Custom Dictionaries" tab.

Click on the "Add Dictionary" button to create a new custom dictionary.

Enter a name for the custom dictionary and select the language of the words you want to add.

Click on the "Add Word" button to add the words to the custom dictionary.

Enter the word and its pronunciation and select the capitalization option as "lowercase".

Once you've added the words to the custom dictionary, the Speech Studio should recognize them without capitalizing them. You can then use the custom dictionary in your Speech Studio project by selecting it in the "Language Model" tab.

Amy Highton 0 Reputation points

2023-11-20T23:05:38.9033333+00:00

This sounds great, but I can't find it! Could you please walk me through where it is, ideally with screenshots? Is it the same as the Custom Keyword feature? If so, how do you tie that to a custom model?
Sign in to comment

Use comments to ask for clarification, additional information, or improvements to the question.

Share via

Audio + transcript failing to export from Editor to Training and Testing Dataset in Azure Speech Studio

1 answer

Your answer