`en-US-GuyNeural not working properly

Question

`en-US-GuyNeural not working properly

Rodrigo 1

For the last few days the voices generated with en-US-GuyNeural sounds weird.
Beside, when try to create voices using phonemes I get this error (only with this voice; it doens happen with en-US-AriaNeural)

Status(StatusCode="FailedPrecondition", Detail="MSTTS_StreamingDumpPhones exception:8000FFFF") websocket error code: 1007

romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-17T14:56:18.297+00:00

@Rodrigo Are you using the audio content creation tool from the speech studio or calling the TTS API directly?
What are the inputs or words used in this case for the error? I tried to use the phonemes with some test words or phrases on speech studio and seems to work fine.

Rodrigo 1

@romungi-MSFT Not sure what you mean. I'm using the JS SDK:

   const synthesizer = new sdk.SpeechSynthesizer(speechConfig, undefined);  
     
       const ssml = `<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xmlns:mstts="https://www.w3.org/2001/mstts" xml:lang="en-US">  
             <voice name="${voiParam.name}">  
             <prosody rate="${voiSpeed}">              
                 ${text}              
               </prosody>  
             </voice>  
             </speak>  
          `;  
         
       synthesizer.speakSsmlAsync(  
         ssml,  
   ....

For phonemes, I'm inserting within the previos ssml:

   <phoneme alphabet="sapi" ph="${el}"></phoneme>

Any phoneme combination raises an error, ie: H AA T

romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-22T07:22:56.803+00:00

@Rodrigo I meant using the audio content creation tool from azure speech studio. I think it might be easier to debug your scenario if you could create the required sound from the tool and then update your SSML to reflect the same with the parameters.
With the current error it is only possible to lookup error details if there is access to your subscription by the speech service team.

romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-17T14:56:18.297+00:00

@Rodrigo Are you using the audio content creation tool from the speech studio or calling the TTS API directly?
What are the inputs or words used in this case for the error? I tried to use the phonemes with some test words or phrases on speech studio and seems to work fine.
Rodrigo 1 Reputation point

2021-02-18T17:27:21.843+00:00

@romungi-MSFT Not sure what you mean. I'm using the JS SDK:

const synthesizer = new sdk.SpeechSynthesizer(speechConfig, undefined); const ssml = `<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xmlns:mstts="https://www.w3.org/2001/mstts" xml:lang="en-US"> <voice name="${voiParam.name}"> <prosody rate="${voiSpeed}"> ${text} </prosody> </voice> </speak> `; synthesizer.speakSsmlAsync( ssml, ....

For phonemes, I'm inserting within the previos ssml:

<phoneme alphabet="sapi" ph="${el}"></phoneme>

Any phoneme combination raises an error, ie: H AA T
romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-22T07:22:56.803+00:00

@Rodrigo I meant using the audio content creation tool from azure speech studio. I think it might be easier to debug your scenario if you could create the required sound from the tool and then update your SSML to reflect the same with the parameters.
With the current error it is only possible to lookup error details if there is access to your subscription by the speech service team.

Answer 1

Rodrigo 1

@romungi-MSFT

ok, understood. Thanks! I didn't even know about Studio Speech.

I got the error. It was because there was no text for the voice side of the phenomenon:

So, this works:

   <voice name="Microsoft Server Speech Text to Speech Voice (en-US, GuyNeural)"><phoneme alphabet="ipa" ph="k aʊ">cow</phoneme></voice></speak>

but this doesn't (but it used to work a few weeks ago)

   <voice name="Microsoft Server Speech Text to Speech Voice (en-US, GuyNeural)"><phoneme alphabet="ipa" ph="k aʊ"></phoneme></voice></speak>

But what I don't understand is that when I do TTS on this word (cow) using the standard way (w/o phonemes) the sound is quite weird.

   <voice name="Microsoft Server Speech Text to Speech Voice (en-US, GuyNeural)">cow</voice></speak>

But if try out Aria, it's the other way around: the standard sounds better than the phoneme.

Is there any reason why this happens?

romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-23T04:44:00.043+00:00

@Rodrigo Usually the neural voices sound much more normal than standard. But there could be some words/numbers which might not sound normal, some edge scenarios. We can ask our team to review these words so they can correct it if there is an issue. Thanks!!
Rodrigo 1 Reputation point

2021-02-25T20:38:20.903+00:00

@romungi-MSFT
Is there any website where to submit issues/bugs with words not sounding natural?
romungi-MSFT 39,696 Reputation points Microsoft Employee

2021-02-26T07:26:44.127+00:00

@Rodrigo Cognitive services user voice forums was used to submit issues similar to this but this site is now inactive. I would request you to post your feedback on this thread for all words/voices that are not accurate and I can reach out internally to help you and pass the feedback. Thanks.

`en-US-GuyNeural not working properly

1 answer

Activity