High-definition voices in Azure Speech

Azure Speech continues to advance text-to-speech technology with neural high-definition (HD) voices. These HD voices understand content, automatically detect emotions in input text, and adjust speaking tone in real-time to match sentiment. They maintain consistent voice personas while delivering enhanced expressiveness, naturalness, and control.

HD voice overview

Azure Speech offers one advanced HD voice model currently:

Key features of HD voices

The following table describes the key features of Azure Speech HD voices:

Comparison of Azure Speech HD voices to other Azure text to speech voices

How do Azure Speech HD voices compare to other Azure text to speech voices? Here's a detailed comparison:

Supported Azure Speech HD voices

Azure Speech provides two sets of HD voices with different model architectures:

Dragon HD voices

The Azure Speech HD voice values use the format voicename:DragonHD:version. The name before the colon, such as en-US-Ava, is the voice persona name and its original locale.

To make sure you use the latest version of the base model that Microsoft provides, use the LatestNeural version.

For example, for the persona en-US-Ava, you can specify:

• en-US-Ava:DragonHDLatestNeural: Always uses the latest version of the DragonHD base model.

The following table lists the available DragonHD voices:

The following styles and paralinguistic tags are supported in HD voices:

Dragon HD Flash voices

HD Flash voices are optimized variants of selected DragonHD voices, currently supporting Chinese (zh-cn) and English (en-US) text. These voices deliver enhanced naturalness and are only available in chinanorth3 currently.

The following table lists all available HD Flash voices and supported styles.

How to use Azure Speech HD voices

Use the same Speech SDK and REST APIs for HD voices as you do for non-HD voices.

Consider these key points when using Azure Speech HD voices:

• Voice locale: The locale in the voice name indicates its original language and region.
• Base models:
  • HD voices include a base model that understands the input text and predicts the speaking pattern accordingly. You can specify the desired model, such as DragonHDLatestNeural, based on the availability of each voice.
• SSML usage: To reference a voice in SSML, use the format voicename:basemodel:version. The name before the colon, such as de-DE-Seraphina, is the voice persona name and its original locale. The base model is tracked by versions in subsequent updates.
• Temperature parameter:
  • The temperature value is a float ranging from 0 to 1, influencing the randomness of the output. You can adjust the temperature parameter to control the variation of outputs. Less randomness yields more stable results, while more randomness offers variety but less consistency.
  • Lower temperature results in less randomness, leading to more predictable outputs. Higher temperature increases randomness, allowing for more diverse outputs. The default temperature is set at 1.0.

Here's an example of how to use Azure Speech HD voices in 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-Ava:DragonHDLatestNeural' parameters='temperature=0.8'>Here is a test</voice>
</speak>

Supported and unsupported SSML elements for Azure Speech HD voices

The Speech Synthesis Markup Language (SSML) with input text determines the structure, content, and other characteristics of the text to speech output. For example, you can use SSML to define a paragraph, a sentence, a break or a pause, or silence. You can wrap text with event tags such as bookmark or viseme that your application processes later.

The Azure Speech HD voices support different SSML elements depending on the model:

• DragonHD voices: Support a subset of SSML elements (see table below)

For detailed information on the supported and unsupported SSML elements for Azure Speech HD voices, refer to the following table. For instructions on how to use SSML elements, refer to the Speech Synthesis Markup Language (SSML) documentation.

Parameter enhancePronunciation

The enhancePronunciation parameter enables enhanced pronunciation handling during speech synthesis. When set to true, the NeuralHD voices apply extra pronunciation optimizations to improve the clarity and correctness of spoken output, particularly for complex, ambiguous, or nonstandard text.

When you enable enhancePronunciation, the service prioritizes pronunciation accuracy by applying enhanced linguistic processing during synthesis. This improvement can help how the system reads:

• Proper nouns, names, and uncommon words
• Acronyms, abbreviations, and mixed-case text
• Words with multiple possible pronunciations depending on context This parameter complements existing pronunciation controls such as SSML-based pronunciation tags and lexicons, and doesn't replace them. The default value is false to preserve predictable, backward-compatible speech output. Enable it when you want the service to apply extra pronunciation optimizations for improved clarity and naturalness.

<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="enhancePronunciation=true">
    This is a pronunciation enhanced example for technical terms like
    Kubernetes, Azure OpenAI, and multilingual content such as 今、何か軽く摘めそうなものある？
  </voice>
</speak>

Recommended use cases

Enable enhancePronunciation in scenarios with structured or technical domain-specific content.

Related content

• Try the text to speech quickstart in Azure Speech
• Learn more about how to use SSML and events