Data, privacy, and security for text to speech

• Recorded voice talent acknowledgement statement file. Customers are required to upload a specific recorded statement spoken by the voice talent in which they acknowledge that you will use their voice to create synthetic voice(s).

  Note

  When preparing your recording script, make sure you include the required acknowledgement statement for the voice talent to record. You can find the statement in multiple languages here. The language of the acknowledgement statement must be the same as the language of the audio recording training data.

• Training data (including audio files and related text transcripts). This includes audio recordings from the voice talent who has agreed to use their voice for model training and the related text transcripts. In a custom neural voice pro project, you can provide your own text transcriptions of audio or use the automated speech recognition transcription feature available within Speech Studio to generate a text transcription of the audio. Both the audio recordings and the text transcription files will be used as voice model training data. In a custom neural voice lite project, you will be asked to record the voice speaking the Microsoft defined script in Speech Studio. Text transcripts are not required for personal voice features.

• Text as the test script. You can upload your own text-based scripts to evaluate and test the quality of the custom neural voice model by generating speech synthesis audio samples. This does not apply to personal voice features.

• Text input for speech synthesis. This is the text you select and send to the text to speech service to generate audio output using your custom neural voice.

• Recorded avatar talent acknowledgement statement file. Customers are required to upload a specific, video-recorded statement spoken by the avatar talent in which they acknowledge that you will use their voice and image to create avatar model(s).

  Note

  When making your video recording plan, make sure you include the required acknowledgement statement for the avatar talent to record. You can find the statement in multiple languages here. The language of the acknowledgement statement must be the same as the language of the video recording training data.

  The acknowledgement statement video is not required to be in the same condition (such as recording environment, avatar talent clothes) as the video recording training data, but the actor’s face should be fully exposed and without any occlusion.

• Training data (including audio files and image/video files). This includes video and audio recordings from the avatar talent who has agreed to use their image and voice for model training. We don’t require a text transcription of the audio in the video.

• Text as the test script. You can upload your own text-based scripts to evaluate and test the quality of the custom avatar model by generating synthetic avatar video samples.

• Text input for speech synthesis. This is the text you select and send to the text to speech service to generate audio and video output using your custom avatar.

Recorded acknowledgement statement verification

Microsoft requires customers to upload an audio file to Speech Studio with a recorded statement of the voice talent acknowledging that the customer will use their voice to create a synthetic voice. Microsoft may use Microsoft’s speech to text and speech recognition technology to transcribe this recorded acknowledgement statement to text and verify that the content in the recording matches the pre-defined script provided by Microsoft. This acknowledgement statement, along with the talent information you provide with the audio, is used to create a voice talent profile. You must associate training data with the relevant voice talent profile when initiating custom neural voice training.

Microsoft may also process biometric voice signatures from the recorded acknowledgement statement file of the voice talent and from randomized audio from the training dataset(s) to confirm that the voice signature in the acknowledgement statement recording and the training data recordings match with reasonable confidence using Azure AI Speaker Verification. A voice signature may also be called a “voice template” or “voiceprint” and is a numeric vector that represents an individual’s voice characteristics that is extracted from audio recordings of a person speaking. This technical safeguard is intended to help prevent misuse of custom neural voice, by, for example, preventing customers from training voice models with audio recordings and using the models to spoof a person’s voice without their knowledge or consent.

The voice signatures are used by Microsoft solely for the purposes of speaker verification or as otherwise necessary to investigate misuse of the services.

The Microsoft Products and Services Data Protection Addendum (“DPA”) sets forth customers’ and Microsoft’s obligations with respect to the processing and security of Customer Data and Personal Data in connection with Azure and is incorporated by reference into customers’ enterprise agreement for Azure services. Microsoft’s data processing in this section is governed under the Legitimate Interest Business Operations section of the Data Protection Addendum.

Training a custom neural voice model

The training data (speech audio) customers submit to Speech Studio is pre-processed using automated tools for quality checking, including data format check, pronunciation scoring, noise detection, script mapping, etc. The training data is then imported to the model training component of the custom voice platform. During the training process, the training data (both voice audio and text transcriptions) are decomposed into fine-grained mappings of voice acoustics and text, such as a sequence of phonemes. Through further complex machine leaning modeling, the service builds a voice model, which then can be used to generate audio that sounds similar to the voice talent and can even be generated in different languages from the training data recording. The voice model is a text to speech computer model that can mimic unique vocal characteristics of a particular speaker. It represents a set of parameters in binary format that is not human readable and does not contain audio recordings.

A customer’s training data is used only to develop that customer’s custom voice models and is not used by Microsoft to train or improve any Microsoft text to speech voice models.

Speech synthesis/audio content generation

Once the voice model is created, you can use it to create audio content through the text to speech service with two different options.

For real time speech synthesis, you send the input text to the text to speech service via the TTS SDK or RESTful API. Text to speech processes the input text and returns output audio content files in real time to the application that made the request.

For asynchronous synthesis of long audio (batch synthesis), you submit the input text files to the text to speech batch service via the Long Audio API to asynchronously create audios longer than 10 minutes (for example audio books or lectures). Unlike synthesis performed using the text to speech API, responses aren't returned in real time with the Long Audio API. Audios are created asynchronously, and you can access and download the synthesized audio files when they are made available from the batch synthesis service.

You can also use your custom voice model to generate audio content through a no-code Audio Content Creation tool, and choose to save your text input or output audio content with the tool in Azure storage.

Data processing for custom neural voice lite (Preview)

Custom neural voice lite is a project type in public preview that allows you to record 20-50 voice samples on Speech Studio and create a lightweight custom neural voice model for demonstration and evaluation purposes. Both the recording script and the testing script are pre-defined by Microsoft. A synthetic voice model you create using custom neural voice lite may be deployed and used more broadly only if you apply for and receive full access to custom neural voice (subject to applicable terms).

The synthetic voice and related audio recording you submit via Speech Studio will automatically be deleted within 90 days unless you gain full access to custom neural voice and choose to deploy the synthetic voice, in which case you will control the duration of its retention. If the voice talent would like to have the synthetic voice and the related audio recordings deleted before 90 days, they can delete them on the portal directly, or contact their enterprise to do so.

In addition, before you can deploy any synthetic voice model created using a custom neural voice lite project, the voice talent must provide an additional recording in which they acknowledge that the synthetic voice will be used for additional purposes beyond demonstration and evaluation.

Data processing for personal voice API (Preview)

Personal voice allows customers to create a synthetic voice using a short human voice sample. The verbal acknowledgement statement file described above is required from each user who uses the integration in your application. Microsoft may process biometric voice signatures from the recorded voice statement file of each user and their recorded training sample (a.k.a the prompt) to confirm that the voice signature in the acknowledgement statement recording and the training data recording matches with reasonable confidence using Azure AI Speaker Verification.

The training sample will be used to create the voice model. The voice model can then be used to generate speech with text input provided to the service via the API, with no additional deployment required.

Video acknowledgement statement verification

Microsoft requires customers to provide a video file with a recorded acknowledgement statement from their avatar talent acknowledging that the customer will use the talent’s image and voice to create a custom avatar. Microsoft will verify that the content in the acknowledgement statement recording matches the pre-defined script provided by Microsoft.

Microsoft may also compare the face of the avatar talent in the recorded acknowledgement video statement file with randomized videos from the relevant training datasets to confirm that the person in the training dataset video recordings and the person in the acknowledgment video statement file match.

Only when both the acknowledgement statement video recording file and video training data are provided and pass actor verification testing will Microsoft start to train the custom avatar model.

Training a custom text to speech avatar model

During the Preview stage of custom text to speech avatar, acknowledgement statement file verification, training data quality check and processing, and custom avatar model training will be performed manually by the text to speech product team, until a customer self-service tool is developed.

After the custom avatar model is trained, customers can access their model on Speech Studio, and select to deploy and use the model at their Azure subscription.

Once the custom avatar model is created, you can use it to create video content through the text to speech service with two different options.

For real time avatar speaking synthesis, you send the input text to the text to speech avatar service via the text to speech avatar SDK. text to speech avatar processes the input text and returns output video stream in real time to your application that made the request.

For asynchronous synthesis (batch synthesis), you submit the input text files to the text to speech avatar batch service via the batch API to asynchronously create video file. Unlike synthesis performed using the streaming API, responses aren't returned in real time with the batch API. Videos are created asynchronously, and you can access and download synthesized videos when it is made available from the batch synthesis service.

You can also use your custom avatar to generate video content through a no-code text to speech avatar tool.

Recorded acknowledgement statement and Speaker Verification data: The voice signatures are used by Microsoft solely for the purposes of speaker verification or as otherwise necessary to investigate misuse of the services. The voice signatures will be retained only for the time necessary to perform speaker verification, which may occur from time to time. Microsoft may require this verification before allowing you to train or retrain custom neural voice models in Speech Studio, or as otherwise necessary. Microsoft will retain the recorded acknowledgement statement file and voice talent profile data for as long as necessary to preserve the security and integrity of Azure AI Speech.

Custom neural voice models: While you maintain the exclusive usage rights to your custom neural voice model, Microsoft may independently retain a copy of custom neural voice models for as long as necessary. Microsoft may use your custom neural voice model for the sole purpose of protecting the security and integrity of Microsoft Azure AI services.

Microsoft will secure and store copies of each voice talent's recorded acknowledgement statement and custom neural voice models with the same high-level security that it uses for its other Azure Services. Learn more at Microsoft Trust Center.

Training data: You submit voice training data of voice talent to generate voice models via Speech Studio, which will be retained and stored by default in Azure storage (See Azure Storage encryption for data at REST for details). You can access and delete any of the training data used to build voice models via Speech Studio.

You can manage storage of your training data via BYOS (Bring Your Own Storage). With this storage method, training data may be accessed only for the purposes of voice model training and will otherwise be stored via BYOS.

Recorded acknowledgement statement data: Microsoft may compare the acknowledgement statement video and training data before allowing you to train or retrain custom avatar models. Microsoft will retain the recorded acknowledgement statement file and avatar talent profile data for as long as necessary to preserve the security and integrity of Azure AI Speech.

Custom text to speech avatar models: While you maintain the exclusive usage rights to your custom avatar model, Microsoft may independently retain a copy of custom avatar models for as long as necessary. Microsoft may use your custom avatar model for the sole purpose of protecting the security and integrity of Microsoft Azure AI services.

Microsoft will secure and store a copy of avatar talent’s recorded acknowledgement statement and custom avatar models with the same high-level security that it uses for its other Azure Services. Learn more at Microsoft Trust Center.

Training data: During the Preview stage, until an automated self-serve training pipeline is available, Microsoft is offering white glove services to train customers’ custom avatar models, with customers’ prior consent. Training data that you provide to Microsoft will be retained and stored by default in Azure storage. Microsoft will delete your training data or avatar model within a reasonable time upon request.

Preventing abuse and harmful content generation (preview)

To reduce the risk of harmful use of the Azure text to speech avatar services, the Azure text to speech avatar includes content safety features. Content safety occurs synchronously as the service processes text to generate audio. No text or generated results are stored in the content classifier models, and text and results are not used to train, retrain, or improve the classifier models. To learn more about the harm categories of content safety features click here.

See also

• Transparency note and use cases for custom neural voice
• Responsible deployment of synthetic speech
• Disclosure for voice and avatar talent
• Disclosure design guidelines
• Disclosure design patterns