Who are you?
We are an AI studio in Seattle. We build text generation and transformation systems that bring content to life with a tone of voice. We also provide consulting services to companies considering text generation or transformation.
Our heritage is in shipping spoken and natural language systems at scale. Our founder Stephen Potter was an early member of the Amazon Alexa AI team (2011-2019) and has launched server and desktop products at Amazon, Microsoft and other companies.
What do you mean by tone of voice?
By tone of voice in text, we mean the lingustic and pragmatic characteristics of a persona or text corpus that demonstrate a particular style of writing or speaking. We are successful when we reflect these characteristics in the content that we generate. A tone of voice can be as general, as in formal vs. casual; or specific, such as 'write like Ernest Hemingway'.
We are also investigating the means to apply the spoken characteristics of a tone of voice to the generation of spoken audio output using text-to-speech (TTS) synthesis systems.
How do you create text with a tone of voice?
We can output text that reflects a given tone of voice in three general ways, depending on the source material used for generation.
We can take raw data such as the codes and numbers in a weather forecast schema, and create a coherent forecast in the manner required:
{day: "1", "forecast": {"code": 800, "high": 63, "low": 42}}
↓
[Regency English model]
↓
I cannot but suppose that tomorrow shall be blessed with the bluest of skies, with a temperature reaching 63 degrees at its utmost and 42 at its lowermost.
This is how our Alexa experience
Miss Jane
works.
We can use prompts combined with a Large Language Model (LLM) such as OpenAI's ChatGPT. This allows for generation based on some preceding context. For example, we can generate an answer given a question; a conversational turn given the preceding conversation; or the next paragraph of a body of text.
[{"role": "system", "content": "Answer the question in the tone of voice of (person)"}, {"role": "user", "content": "What car do you drive?"}]
↓
[(person='Cardi B') + LLM]
↓
Girl, I got a few cars but my favorite is my Lamborghini Aventador! It's bright orange, just like my personality, haha!
Our Alexa experience The Chat Room uses this approach
We can take existing text in any form, and revoice the content, maintaining the semantics of the original, but giving it a new tone of voice. This is sometimes called style transfer.
Who do you want to talk to?
↓
['high-formality' transformation model]
↓
To whom do you wish to speak?
We'd love to hear from you!
If you have a scenario for generating or transforming text into a certain tone of voice, or if you have feedback on any of our products, please get in touch.
