Quality of live chat avatar synthesis (azure speech services)

Question

Quality of live chat avatar synthesis (azure speech services)

Axel Angel 20

Hello, I'm looking for ways to improve the quality and the natural feeling of the real-time avatar synthesis: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech-avatar/what-is-text-to-speech-avatar

Are there options passed to the API to improve the quality of the video feed or make it more natural?
Are there avatars with different ethnicity? We found all of them exhibit Asian traits. I'm looking for avatars that match depending on the region of the user.
If the option below cannot be solved I'm considering making custom avatars, however some points aren't clean it he documentation: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech-avatar/custom-avatar-record-video-samples
- Once we have all material video/audio for training, how much time does it take from the time we submit the material until we can use it in the API directly? Including submission process, and training.
- What is the pricing of those steps: how much does it cost to train an avatar audio+voice (the compute time cost) and how much does it cost to use it in production (inference generation).

The userbase has criticism of the current quality thus the need for improvements. Can you help me? I'm looking for ways to improve the quality however I'm considering alternatives because other services and new ones seem to offer very high quality output. My hope is to make azure generation better with your help. Thank you.

0 comments

Answer accepted by question author

0 additional answers

Your answer

Answer 1

Hi Axel Angel

Yes, Custom text to speech Avatar would address your requirement here.

Here are answers to you queries

Once we have all material video/audio for training, how much time does it take from the time we submit the material until we can use it in the API directly? Including submission process, and training.

You need to take approval first for Custom Avatar usage here through this form

Then create high quality training video as per guideline for Custom Avatar Training.

Reference- https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech-avatar/custom-avatar-record-video-samples

Once you have finetuned model Avatar, it will be usable through SDK or UI.

Training for custom audio takes up to 40 hrs.

Reference - https://learn.microsoft.com/en-us/azure/ai-services/speech-service/custom-neural-voice

Training for custom Avatar depends on training file size and complexity (20-48 hrs)

What is the pricing of those steps: how much does it cost to train an avatar audio+voice (the compute time cost) and how much does it cost to use it in production (inference generation).

Compute training cost for Avatar training

Around 52$ dollar per training compute hour and 5$ per hour for endpoint hosting per 2024 data

Please check with Sales team for exact data in 2025.

The Avatar service is billed at approximately $1.44 per 6-minute block of generated video output, while the neural TTS is billed separately at around $4.00 per 1 million characters of synthesized speech.

So, both services contribute to the total cost: TTS for generating the speech, and the Avatar service for rendering the video with lip-sync and facial expressions.

Reference

https://learn.microsoft.com/en-us/answers/questions/3207134/unable-to-estimate-avatar-usage

https://learn.microsoft.com/en-us/answers/questions/4372884/struggling-to-calculate-costs-for-a-few-services-f

https://azure.microsoft.com/en-us/pricing/details/cognitive-services/speech-services/

Improving Quality of default avatars.

Utilize Speech Synthesis Markup Language (SSML). SSML allows you to fine-tune aspects like pitch, pronunciation, speaking rate, and volume for the text-to-speech output. This can help in producing more natural-sounding speech. You can learn more about SSML here.

Please accept this answer if it helped.

Hope it helps.

Thank you

Axel Angel 20 Reputation points

2025-11-06T14:50:37.3366667+00:00

Thank you for your thorough answer with detailed explanation and your estimated cost. This is very helpful. I need time to properly assess the next step on my side.

One thing that wasn't addressed is the question of avatar variety:

Are there avatars with different ethnicity? We found all of them exhibit Asian traits. I'm looking for avatars that match depending on the region of the user.

Maybe I'm missing a part where I have more choices here. That would help to know, will this in the near future? Are there other solutions except for the custom avatar maybe? I'm crurently investigating our possibilities here and collecting information. Thank you for help.
Manas R Mohanty 17,270 Reputation points Moderator

2025-11-06T16:02:54.6433333+00:00

Hi Axel Angel

The standard avatars provided by Azure AI Speech are limited in diversity. Most of the prebuilt avatars currently available do have an Asian appearance. This has been confirmed in Microsoft Q&A discussions, where users noted the lack of South Asian or other ethnic representations. [learn.microsoft.com]

There is no automatic region-based avatar selection. Even if you choose a different Azure region for your Speech resource (e.g., East US, West Europe, Southeast Asia), the avatar appearance does not change based on geography. Regions affect data residency and API endpoints availability, not avatar visuals.

Thank you.