Something changed with the zh-CN-XiaoxiaoNeural voice - the 'sentiment' expression no longer works, as of a few days ago. Using exactly the same code, but the output is quite different (no longer sounds sentimental). Was an update posted? If so, where do we find a list of changes?

For West Europe region, the service is returning "Specified grammar type is not supported!" when we pass in LUIS reference grammar ID (a.k.a. IntentRecognizer).

This causes the speech service to fail with a "WebSocket is already in CLOSING or CLOSED state." error when the LUIS reference grammar ID is passed in. If it is not included, the service works correctly.

May be related to this issue in the Cognitive Services Speech SDK repo: https://github.com/Azure-Samples/cognitive-services-speech-sdk/issues/127

The custom speech has good accuracy compared to others solutions but it doen't support pronunciation for portuguese BR, so important words for business aren't recognized. It's a very important feature for business usage. The UI offers only support for english language.

At present, we have to determine the locale of the input language in detail, such as "Speech to Text" feature. For example, en-US, en-AU, and so on.
Users may not know which one to choose, so it will be easier to use if they automatically recognize it from their voice.

Please let me know if you have any prospects for the future.

Norwegian needs a grammatical rethink with regards to compound words.

Currently in Norwegian we distinguish individual words by putting spaces between them, so for instance when I mention to the speech-to-text-service the console window ("konsollvindu" in Norwegian), it outputs it as "konsoll vindu" with a space between the words.

This.

Must.

Absolutely.

Be.

Fixed.

I am a linguist by degree. Please hire me to fix this if you need help, because you really do.

Depending on screen resizing, certain elements on cris.ai will not be visible. There seems to be an issue with the underlying architecture of the site since this is the second issue that arose from resizing the window. Both times, things went wrong when the window occupied half or less than half of the monitor screen.

We are currently using Bing Speech with LUIS, but looking to convert to Speech service.

Right now we have multiple recorders that operate in the browser, Flash, WebRTC, HTML5.

Each of these has to connect to Bing Speech to Text to get realtime translation and LUIS results to drive actions in the application. Additionally we are currently streaming the audio to Amazon S3. Ideally we would like to stream the audio only once, and have it picked up by Microsoft from Speech to Text AND be able to retrieve a URL for later use.

Hi,

The new neural text 2 speech feature looks amazing, but one language is missing : French :)

I don't know if it something on the list, or if it is coming soon, but I'm waiting for this feature to switch from Google Wavenet.
I'm pretty sure that the french voice generated by this new neural TTS by MS Cognitive service will be a game changer.

The french language is complex, the emphasis, the punctation etc... but if MS can provide the same awesome quality as they have done in english... we will be able to build something incredible...

The lack of a confidence score on word level feature is a show stopper for my company's project. It would be extremely useful for us to have the confidence score included within "Words" list , which consist of words and their timestamps.

According to this answer: https://social.msdn.microsoft.com/Forums/en-US/4979ca92-aa0f-4d09-b010-fc2eeb1bde80/speech-results-confidence-score-on-word-level?forum=AzureCognitiveService#8ae67445-4e23-49ea-b694-a8d877dc2dd0
the feature is not public and we suspect that it could be provided quickly.

I'd be grateful for each vote for this idea!

Please add Nigerian Yoruba language or give opportunity to specific users to develop a model with local language.
Thank you

Once the synthesizer starts synthesizing audio, it can't be stopped. It would be nice to have some method that would stop/interrupt currently running audio synthesis.

For training a model, I only have access to use "related text" and not audio+transcription.

When training a model on speech.microsoft.com fails we need proper error messaging. Right now it literally says "Failed" which provides absolutely no useful information for debugging.

Why do you hate your support people that they have to deal with customers whose only info is "it failed"...

Note that this is happening with training data that has already successfully been uploadted.

Add support for wake word detection on IOS platform

Support more languages/derivatives for Custom Voice Fonts. In particular, I'm looking to create a custom voice for en-GB, and building it using en-US doesn't quite work. I would also be looking to create custom voices for other flavours of English, en-AU being a priority.

Azure Speech API should offer json mark labels for Text to Speech audio. This allows developers to use the audio file and the json mark labels to create audio tracking text in the app. The competitor has a similar solution. I found Azure TTS to be superior but am forced to use the competitor's solution due to lack of json mark labels. Speech mark labels should be in json format and available for all languages. It should provide information such as the begin and end timestamp of each sound to the text, phrase and sentences.

Add support for other audio formats and bitrates

Please add Gaeilge the Irish language.