Lots of text-to-speech service providers use AI voices from Amazon Web Services, Yandex, Microsoft Azure, and Google Cloud Platform.

But these voices sound best when you use them in BeyondWords.*

Why?

It's thanks to our natural language processing algorithms.

Setting voices up for success

AI voices can interpret text in two formats: plain text, or speech synthesis markup language (SSML).

SSML tags provide extra information to the AI voice, clarifying pronunciations and improving speech flow. Using SSML therefore ensures a higher-quality voice output.

Some text-to-speech services only allow you to input plain text, meaning you can't achieve higher-quality outputs through SSML.

Others, like Amazon Polly, give you the option to manually insert SSML tags. But this is complex and time-consuming. Let's say the voice is mispronouncing "Joe Biden". Fixing this requires an understanding not only of SSML, but of the international phonetic alphabet (IPA) - symbols that linguists use to represent speech sounds.

Joe Biden

This is not feasible for the majority of publishers.

BeyondWords, on the other hand, adds the SSML tags for you.

Whether it is imported from your website in HTML format or manually imported as plain text, your content is automatically converted into SSML before being processed by the AI voice.

This is made possible by a layer of natural language processing (NLP) algorithms, which can programmatically read, analyze, and interpret written language.

How our NLP works

Our NLP uses a combination of rule-based and neural network-based techniques. Our deep learning models are trained on large data sets, which allow them to "learn" how humans convert particular text elements into speech and how this differs depending on context. This is particularly useful for resolving ambiguities in text.

For example, in the sentence "I read the book", our NLP identifies, through contextual features, that the homograph "read" is most likely being used in the past tense, and so should be pronounced like "red" (/rɛd/) as opposed to "reed" (/riːd/). It applies the SSML tag accordingly, ensuring the AI voice gives the correct output.

I read the book yesterday.

Also consider non-standard elements, such as dates. Our system can determine whether to read a number as cardinal (e.g. twenty nine) or ordinal (e.g. twenty ninth) numeral based on the usage context and other features.

Without this NLP and SSML layer, the AI voice may not predict which pronunciation of an element is correct and simply output a naive "best guess". This is why many text-to-speech systems struggle with ambiguities.

A customizable and evolving NLP

Our team of computational linguists make iterative and domain-specific improvements to the NLP. This means that voice outputs evolve even when the AI voices behind them stay the same, and can adapt to the needs of each BeyondWords user.

Take the abbreviation "NLP", for instance. In the context of this article, it refers to "natural language processing". But it is also the airport code for Nelspruit, South Africa. In medicine, it can be shorthand for 'no light perception'. With custom-built text normalization rules, we can ensure that our system delivers the most relevant result for a particular publisher.

Text-to-speech providers without a post-optimization NLP layer have no way to efficiently extend or make domain-specific customizations to an existing voice. If they wish to correct conversion errors, they must retrain the voice itself - something that comes at great cost and cannot guarantee accuracy, especially when it comes to unusual and idiosyncratic text-to-speech conversions.

Here are some more examples of what our NLP can do:

Apply

SSML tags to ensure correct pronunciation of novel or complex words

Apply

SSML tags to ensure foreign words are pronounced accurately

Apply

SSML tags to ensure s...