Filler words sometimes cause issues with wordBoundary results

[dwegscheidTSC]

I'm using the microsoft-cognitiveservices-speech-sdk@1.28.0 on Node.js

When using some voices (en-US-SaraNeural does this the most frequently for me, but I've run across it with other voices), filler words like "um" or "uh" cause the wordBoundary event arguments to get garbled after the filler word. Both the text and timing don't match the audio when this happens.

Example code

Here's a small example that illustrates what I'm running into for the text becoming inaccurate:

synthesizeSpeech: async (): Promise<string[]> => {
    const speechConfig = SpeechConfig.fromSubscription(key, region);
    speechConfig.speechSynthesisOutputFormat = SpeechSynthesisOutputFormat.Audio48Khz192KBitRateMonoMp3;

    const synthesizer = new SpeechSynthesizer(speechConfig, null as unknown as AudioConfig);
    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="en-US-SaraNeural">
             Hi my name is uh Sara. I am a voice that gives good audio but incorrect word boundary results. 
          </voice>
       </speak>
    `;

    return new Promise<string[]>((resolve, reject) => {
       const onError = (err: string | Error): void => {
          synthesizer.close();
          reject(err);
       };

       const words: string[] = [];
       synthesizer.wordBoundary = (_, event) => {
          words.push(event.text);
       };
       synthesizer.speakSsmlAsync(ssml, result => {
          if (result.errorDetails) {
             onError(result.errorDetails);
          } else if (result.reason === ResultReason.SynthesizingAudioCompleted) {
             synthesizer.close();
             resolve(words);
          }
       }, onError);
    });
 }

so that await synthesizeSpeech() returns

[
   "Hi",
   "my",
   "name",
   "is",
   "rrec", // everything after where "uh" should be is garbled
   "t",
   "w",
   "rd",
   "b",
   "undar",
   " res",
   "lts. ",
   "    ",
   "     ",
   " </",
   "oice>\n   ",
   "    ",
   "</speak>",
   "",
   ""
]

[yulin-li]

Thanks for reporting this issue. I will forward this to the service experts for further investigation