Text to Speech and Avatars

This content was generated with the assistance of AI and should be verified against the official documentation before being relied on in production.

See the section bibliography for the reference material consulted while preparing these pages.

This page targets Azure CLI 2.90.0, Speech SDK 1.52.0 and the batch synthesis REST API 2024-04-01, checked on 2026-10-06. Voice names and the SSML support matrix change often. Check the voice gallery before you hard-code a voice.

Azure Speech in Foundry Tools also does the reverse of speech to text: it turns text, or SSML markup, into natural speech. The vendor-neutral comparison with other TTS providers is on Text to Speech.

Voices

Voice family Notes

Voice family	Notes
Standard neural voices	Hundreds of voices in more than 140 languages and variants, at 24 kHz (some at 48 kHz). Names look like `en-US-AvaNeural`. Many support speaking styles (`cheerful`, `newscast`, `whispering`) and roles.
HD voices	Higher-quality voices that adapt tone to the text, for example `en-US-Ava:DragonHDLatestNeural`. They support a smaller subset of SSML.
Azure OpenAI voices	OpenAI voices (`alloy`, `nova`, `shimmer` and others) served through Speech, for example `en-US-AlloyMultilingualNeural`.
Multilingual voices	One voice that speaks many languages with the same timbre and switches language inside one text. Useful for mixed-language content such as book titles in their original language.

Standard neural voices

Hundreds of voices in more than 140 languages and variants, at 24 kHz (some at 48 kHz). Names look like en-US-AvaNeural. Many support speaking styles (cheerful, newscast, whispering) and roles.

HD voices

Higher-quality voices that adapt tone to the text, for example en-US-Ava:DragonHDLatestNeural. They support a smaller subset of SSML.

Azure OpenAI voices

OpenAI voices (alloy, nova, shimmer and others) served through Speech, for example en-US-AlloyMultilingualNeural.

Multilingual voices

One voice that speaks many languages with the same timbre and switches language inside one text. Useful for mixed-language content such as book titles in their original language.

SSML

Speech Synthesis Markup Language controls pronunciation and delivery:

<prosody> sets rate, pitch and volume; <break> inserts pauses; <emphasis> stresses words
<mstts:express-as style="..." role="..."> picks a speaking style or role (a Microsoft extension)
<lexicon uri="..."> points to a custom pronunciation lexicon, for example for author names
<phoneme> and <say-as> control how one word, a date or a number is read
<mstts:viseme type="FacialExpression"/> asks for viseme (mouth-shape) events, used to animate an avatar or a 3D character in sync with the audio

Not every voice supports every element. HD and OpenAI voices ignore styles and some prosody settings. Check the SSML support matrix per voice family.

<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-AvaMultilingualNeural">
    <mstts:express-as style="narration-relaxed">
      Chapter one. <break time="600ms"/>
      <prosody rate="-5%">It was a bright cold day in April,</prosody>
      and the clocks were striking thirteen.
    </mstts:express-as>
  </voice>
</speak>

Source: Speech Synthesis Markup Language overview

Provisioning

In the Azure portal

Use the Foundry resource foundry-bookshelf (Resources, Endpoints and Authentication).
In the Foundry portal (https://ai.azure.com), open Playgrounds → Speech playground → Text to speech to try voices and SSML. The Voice gallery lists every voice with samples.
For long content, the Audio Content Creation tool in Speech Studio (https://speech.microsoft.com) is a no-code SSML editor.
Assign Cognitive Services Speech User to the identity that calls the API.

With the Azure CLI

TOKEN=$(az account get-access-token --resource https://cognitiveservices.azure.com --query accessToken -o tsv)
curl -s "https://$FOUNDRY.cognitiveservices.azure.com/tts/cognitiveservices/voices/list" \
  -H "Authorization: Bearer $TOKEN" | jq -r '.[] | select(.Locale=="en-US") | .ShortName' | head

Source: Text to speech REST API — list voices

Synthesis with the Speech SDK

Synthesize a book description to an MP3 file:

import azure.cognitiveservices.speech as speechsdk
from azure.identity import DefaultAzureCredential

speech_config = speechsdk.SpeechConfig(
    endpoint="https://foundry-bookshelf.cognitiveservices.azure.com/",
    token_credential=DefaultAzureCredential(),
)
speech_config.speech_synthesis_voice_name = "en-US-AvaMultilingualNeural"
speech_config.set_speech_synthesis_output_format(
    speechsdk.SpeechSynthesisOutputFormat.Audio24Khz96KBitRateMonoMp3
)
audio_config = speechsdk.audio.AudioOutputConfig(filename="isbn-9780143111597.mp3")
synth = speechsdk.SpeechSynthesizer(speech_config=speech_config, audio_config=audio_config)

result = synth.speak_text_async("A sweeping novel about memory and loss.").get()
assert result.reason == speechsdk.ResultReason.SynthesizingAudioCompleted

Source: How to synthesize speech (Python)

Synthesize SSML from a file and stream the audio to memory (for example to return it from catalog-api):

using Azure.Identity;
using Microsoft.CognitiveServices.Speech;

var config = SpeechConfig.FromEndpoint(
    new Uri("https://foundry-bookshelf.cognitiveservices.azure.com/"), new DefaultAzureCredential());
config.SetSpeechSynthesisOutputFormat(SpeechSynthesisOutputFormat.Audio24Khz96KBitRateMonoMp3);

using var synthesizer = new SpeechSynthesizer(config, null); // null = no speaker, keep audio in memory
string ssml = await File.ReadAllTextAsync("sample-chapter.ssml");
SpeechSynthesisResult result = await synthesizer.SpeakSsmlAsync(ssml);

if (result.Reason == ResultReason.SynthesizingAudioCompleted)
{
    await File.WriteAllBytesAsync("sample-chapter.mp3", result.AudioData);
}

Source: How to synthesize speech (C#)

import com.azure.identity.DefaultAzureCredentialBuilder;
import com.microsoft.cognitiveservices.speech.*;
import java.net.URI;
import java.nio.file.*;

SpeechConfig config = SpeechConfig.fromEndpoint(
    new URI("https://foundry-bookshelf.cognitiveservices.azure.com/"),
    new DefaultAzureCredentialBuilder().build());
config.setSpeechSynthesisVoiceName("en-US-AvaMultilingualNeural");
config.setSpeechSynthesisOutputFormat(SpeechSynthesisOutputFormat.Audio24Khz96KBitRateMonoMp3);

try (SpeechSynthesizer synth = new SpeechSynthesizer(config, null)) {
    SpeechSynthesisResult result = synth.speakText("A sweeping novel about memory and loss.");
    Files.write(Path.of("description.mp3"), result.getAudioData());
}

Source: How to synthesize speech (Java)

To get viseme events, subscribe to viseme_received (Python) or VisemeReceived (C#). Each event has an audio offset and a viseme ID, or a blend-shapes frame when you use <mstts:viseme type="FacialExpression"/>.

Batch synthesis

Real-time synthesis is limited to 10 minutes of audio per request. For longer content, such as a full sample chapter, use the asynchronous batch synthesis API. It accepts plain text or SSML, and writes the audio to a storage location you can download from.

curl -s -X PUT \
  "https://$FOUNDRY.cognitiveservices.azure.com/texttospeech/batchsyntheses/sample-ch1-9780143111597?api-version=2024-04-01" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{
    "description": "Bookshelf audiobook sample, chapter 1",
    "inputKind": "SSML",
    "inputs": [{"content": "<speak version=\"1.0\" xmlns=\"http://www.w3.org/2001/10/synthesis\" xml:lang=\"en-US\"><voice name=\"en-US-AvaMultilingualNeural\">Chapter one...</voice></speak>"}],
    "properties": {
      "outputFormat": "audio-24khz-96kbitrate-mono-mp3",
      "wordBoundaryEnabled": true,
      "concatenateResult": true,
      "destinationContainerUrl": "https://stbookshelf.blob.core.windows.net/media"
    }
  }'

curl -s "https://$FOUNDRY.cognitiveservices.azure.com/texttospeech/batchsyntheses/sample-ch1-9780143111597?api-version=2024-04-01" \
  -H "Authorization: Bearer $TOKEN" | jq '{status, outputs}'

Source: Batch synthesis API

Custom voice

Custom voice builds a voice from recordings of a real person. Both options are Limited Access: you apply, describe your use case, and record the voice talent’s consent statement in the service.

Professional voice fine-tuning trains a brand voice from hundreds to thousands of studio-quality utterances. Training and endpoint hosting are billed separately from synthesis.
Personal voice creates a voice from a short (about 1 minute) consented sample, for apps where users speak with their own voice, for example in translated dubbing.

Source: Custom voice overview

Text-to-speech avatars

An avatar is a photorealistic talking-head video driven by text to speech.

Batch avatar synthesis renders a video file (for example a training clip) asynchronously.
Real-time avatar streams video over WebRTC for interactive agents, and pairs with Voice Live.
Standard avatars are prebuilt characters. Custom avatars are trained from video of a real actor, and are Limited Access with recorded consent.

Source: What is text to speech avatar?

Use cases

Scenario Feature Why this feature Example input → output

Scenario	Feature	Why this feature	Example input → output
Bookshelf: audiobook samples	Batch synthesis with SSML and a multilingual voice	Chapters run over 10 minutes; SSML adds pauses and narration style.	Chapter 1 text → `media/sample-ch1-<isbn>.mp3`, linked from the catalog page
Bookshelf: read-aloud of book descriptions	SDK synthesis to a stream	Short text, needed on demand; cache the MP3 in Blob behind Front Door.	Description → 20-second MP3 played in the browser
Accessibility read-aloud	Neural voices, or Immersive Reader for a full reading UI	Natural voice, many languages.	Article → audio for screen-reader users
IVR prompts	SSML with styles and `say-as`	Correct reading of dates, amounts and order numbers.	"Your order 4512 ships on 7 October" → telephony WAV at 8 kHz
E-learning narration	Batch synthesis, avatar video	Long scripts; consistent voice across modules.	Course script → narrated video
In-car navigation	Neural TTS container	Works offline; low latency.	Route instruction → spoken prompt

Bookshelf: audiobook samples

Batch synthesis with SSML and a multilingual voice

Chapters run over 10 minutes; SSML adds pauses and narration style.

Chapter 1 text → media/sample-ch1-<isbn>.mp3, linked from the catalog page

Bookshelf: read-aloud of book descriptions

SDK synthesis to a stream

Short text, needed on demand; cache the MP3 in Blob behind Front Door.

Description → 20-second MP3 played in the browser

Accessibility read-aloud

Neural voices, or Immersive Reader for a full reading UI

Natural voice, many languages.

Article → audio for screen-reader users

IVR prompts

SSML with styles and say-as

Correct reading of dates, amounts and order numbers.

"Your order 4512 ships on 7 October" → telephony WAV at 8 kHz

E-learning narration

Batch synthesis, avatar video

Long scripts; consistent voice across modules.

Course script → narrated video

In-car navigation

Neural TTS container

Works offline; low latency.

Route instruction → spoken prompt

Bookshelf’s read-aloud page also offers Immersive Reader for readers who want word highlighting and translation.

When not to use it

For a full voice conversation (listen, reason, respond), Voice Live handles synthesis for you with turn-taking built in.
When you need one fixed sentence in a few languages, pre-render the audio once and serve it from Blob, rather than synthesizing on every request.
For expressive, open-ended speech generation, an LLM audio model may fit better. See Choosing a Foundry Tool (or a Foundry Model).

Clean up

curl -s -X DELETE \
  "https://$FOUNDRY.cognitiveservices.azure.com/texttospeech/batchsyntheses/sample-ch1-9780143111597?api-version=2024-04-01" \
  -H "Authorization: Bearer $TOKEN"