Apple looking to add character to text-to-speech voices
An Apple patent application discovered on Thursday outlines an invention that uses metadata from emails, texts and other communications to determine how a synthesized voice sounds in a text-to-speech (TTS) system.
The filing, titled "Voice assignment for text-to-speech output," looks to create "speaker profiles" which can change the voice characteristics of TTS output to match parsed-out metadata like age, sex, dialect and other variables.
As noted by the application, many systems exist today to aid the visually impaired, including the system on Apple's iPhone, however most TTS engines "generate synthesized speech having voice characteristics of either a male speaker or a female speaker. Regardless of the gender of the speaker, the same voice is used for all text-to-speech conversion regardless of the source of the text being converted." Apple's invention proposes a different solution.
Instead of hearing the same voice for every message, the invention obtains metadata "directly from the communication or from a secondary source identified by the directly obtained metadata" to create the most suitable speaker profile.
According to the patent filing, "Providing a speech output that is associated with a speaker profile allows speaker recognition while providing a more enjoyable and entertaining experience for the listener."
An example is provided in which a user receives a message from "Charles Prince," who has an email address of [email protected], regarding a party for "Albert." In this case, the system could use the ".uk" address as primary metadata. Secondary metadata can be gathered if a contact card is attached to the message, or if Charles Prince's information is already in the user's address book.
The data from the text and the corresponding metadata are then fed into a TTS engine, which assigns a speaker profile to convert the text into speech.
After converting each word and phonetic transcription in the text to distinct sounds that comprise a given language, the TTS engine then divides and marks rhythmic sounds like phrases, clauses and sentences.
In some implementations, speech can be created by piecing together pre-recorded voice fragments, including sounds, entire words or even sentences, that are stored on a mobile device or in an off-site database.
In other implementations, the TTS engine can include a synthesizer that "incorporates a model of the human vocal tract or other human voice characteristics to create a synthetic speech output according to the speaker profile."
One of the most interesting iterations notes that "a speaker's voice can be recorded and analyzed to generate voice data."
From the patent filing's description:
For example, the speaker's voice can be recorded by a recording application running on the device or during a telephone call (with permission). The voice characteristics of the speaker can be obtained using known voice recognition techniques. In this implementation, a speaker profile may not be necessary as the speaker's name can be directly associated with voice data stored in voice database.
As for output, the system may pick the ".uk" email address to use as primary metadata, taking contact card information like a birthday to determine sex and age, to subsequently output a speaker profile matching an older male with a British accent. Charles Prince's physical address, phone number, or picture can also be used to determine a speaker profile. The more metadata available, the more refined the output.
It is unclear if Apple plans to deploy such a system, however the company currently has a similar, albeit less advanced, system in place with Siri. While the feature is limited to certain regions, Siri has an option to choose dialects like "English (United States)" or "English (United Kingdom)" to recognize incoming voice commands, as well as provide responses in the selected accent.