Various initials are mispronounced as abbreviations in Dutch by text-to-speech. Since Teams uses this service for the voicemail messages, this is very annoying. These are a few examples:

b. - pronounced as bruto
c. - pronounced as cent
e. - pronounced as edidiet
f. - pronounced as forte
H. - pronounced as heilige
i. - pronounced as in
l. - pronounced as loco
p. - pronounced as piano
t. - pronounced as temperatuur
H.M. - pronounced as Hare Majesteit (Her Majesty)
V.J. - pronounced as vorig jaar (last year).
G.B. - Gedaan en bieden
B.G. - Bovengenoemde

We would like to see that these abbreviations are pronounced normally; letter by letter.

I also made a Uservoice at https://microsoftteams.uservoice.com/forums/908686-bug-reports/suggestions/41672296-voicemail-greeting-don-t-pronounce-initials-as-a

Hi, my company has been using Azure's speech to text service and are happy with the results.

However, we've hit a snag. For the RecordingsUrl parameter, I understand we will be passing in a blob uri that's public facing (no auth required), but our audio files are stored in a 3rd party service whereby an auth bearer token is required to be passed into the header to be able to access the .wav file (request header: { Authorization: Bearer <token> }).

For the speech to text Rest API, is it possible if we can have an additional parameter (eg. RecordingsUrlToken) that we can pass a token to so that it can then use said token to get the .wav file?

Thanks in advance.

I used the following workflow for training my intent recognition:
1) I've a series of entities, features and patterns edited
2) I've a series of example inputs for training
3) all the samples have the entities marked
4) now I train the examples
5) execute a series of batch test cases

The issues recommended for improvement
- The test cases for the batch testing require character positions, startPos and endPos. I had no other option than counting these manually, which is error prone.
- When loading the batch test cases, the feedback / error log is hard to find and hardly descriptive. References are to character numbers. In a 4kByte file counting characters is no fun.
- When I want to make changes to the test cases, I have to edit the file describing all the test cases, delete the original set, load the edited set and provide again the name for the test set. This sure can be optimized.
- When improving my recognition definitions and training set based on the test results, it is advantageous to have access to the test results and the training set editing panes simultaneously.

Considering all the above, I'd like to suggest creating an editing facility for test patterns in the portal itself, reusing the editor of the training set. One could even think of having a generic editor with the option to mark utterances for inclusion in the training set and/or the test set.

It would be nice if you could specify an “extern id” when you create a batch transcription request and that the “external id” is also returned in the response of the web hook callback.

Why? To be able to link a request to an id of a running process/workflow. For example in a durable function. The durable function (using a orchestration) looks like:
1. Durable function send a message to the speech to text service (STTS) to create a transcription.
2. Durable function makes a call to context.WaitForExternalEvent<string>("TranscriptionCompleted");
3. At some point in time the STTS is finished and calls the web hook.
4. The web hook gets the message, but it only contains the transcription id, so it has no knowledge to trigger the correct durable function (sendEventUrl), because to be able to send the TranscriptionCompleted event you need to know the instance id of the durable function.

Of course there are way to solve this “problem” now, for example by creating a mapping table between the transcription id and the durable function orchestration instance id. However this could be potential an a problem is you want to listen for the TranscriptionCreation event, because that could potential be fired before a mapping is created in the mapping table (because you first need to create the request before you can store the transcription id).

So a nice solution for this, is to be able to specify an external id.

This textbook (Computer Security Handbook by Seymour Bosworth et al., Chapter 33) seems to cause countless errors using the Aria Neural voice. (attached txt and pdf were trimmed to respect the copyright of the author)

It messes up the chapter markers, saying "January First, Thirty Three" when it says "33.1.1" (as well as all the other section markers)

802.11 is pronounced "eight hundred and two point one one"

SSIDs is pronounced "sids"

BSSIDs is pronounced "bsids"

2Mb/s (as well as other Mb/s numbers) is pronounced "two em bee slash ess" which should be pronounced 2 Megabits per second.

LAN is pronounced lon instead of lan

VLAN should be pronounced "V-lan" instead of being spelled out or occasionally pronounced "vlan" as in vlog and pan put together.

There are other mistakes in this paper as well, but I don't have time to list them all here. Still, the TTS is in serious need of learning tech vocabulary.