Azure Speech DragonHDLatestNeural produces electrical/static noise during speech in both raw MP3 and RIFF WAV outputs

1. The provided information does not document any known issue where DragonHDLatestNeural voices (such as en-US-Steffan:DragonHDLatestNeural) consistently produce electrical/static noise in the synthesized audio.
2. The documentation describes DragonHDLatestNeural voices as high‑definition neural voices based on large language models, with enhanced contextual understanding and expressive capabilities, but it does not mention static/noise artifacts as an expected or known behavior for specific text or SSML patterns.
3. Supported usage and configuration:
   • DragonHDLatestNeural voices are used through the standard Azure Speech SDK and REST APIs, with the same output formats as non‑HD voices.
   • HD voices can be referenced in SSML using the voicename:basemodel:version format, for example:

          <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-Ava:DragonHDLatestNeural' parameters='temperature=0.8'>Here is a test</voice>
          </speak>
     

   • The documentation lists supported SSML elements for DragonHD voices (for example, <phoneme>, <say-as>, <sub>, <p>, <s>), but does not associate any of them with noise/static issues.
   • There is no guidance in the provided material that recommends specific output formats, regions, or voice parameters (such as temperature, top_p, top_k, or cfg_scale) as a way to avoid static or encoding artifacts. Parameter tuning is documented only for controlling expressiveness, stability, and style for DragonHDOmni voices, not for mitigating noise.
4. Diagnosing model vs. encoding issues:
   • The context does not provide a diagnostic procedure specific to distinguishing between voice‑model artifacts and service‑side encoding problems for Azure Speech HD voices.
   • It does, however, emphasize that HD voices are standard Azure Speech voices that share the same SDK/REST surface and that they support multiple languages and advanced prosody, without calling out any special handling or known defects in the audio encoding pipeline.

Within the supplied documentation, there is no statement that:

• DragonHDLatestNeural voices are known to generate static/electrical noise; or
• Certain SSML patterns or text inputs are known to cause such artifacts; or
• Particular formats/regions/settings are recommended specifically to avoid static.

References:

• High-Definition voices in Azure Speech
• Language and voice support for Azure Speech