Text-to-speech (TTS) technology has come a long way in recent decades. We are no longer talking about robotic, monotone voices that are barely intelligible, but voices that are almost indistinguishable from those of a real human being. And now, with the use of artificial intelligence, voices can even express emotions and speak in multiple languages.
In this article, we will explore some of the best TTS artificial intelligences available today, from the highly acclaimed NaturalSpeech2 to the fun and powerful AudioLM. We will discover how these technologies are changing the way we hear and understand language and how they are helping people with visual and speech disabilities.
Moreover, this article will not be limited to being a boring technical report. We’ll add a touch of humor along the way, because, after all, who said technology had to be boring?
NaturalSpeech2: The most realistic speech
NaturalSpeech2 is a highly advanced TTS model that has received numerous positive reviews for its realistic voice quality. Although the code is not public, the demo is available for use on their website.
The demo is impressive. The voice generated by NaturalSpeech2 is so realistic that, if you listen to it with your eyes closed, you might think it is a real person speaking. The model is capable of expressing a wide variety of emotions, including happiness, sadness, surprise, fear and more.
You can listen to several demos at https://speechresearch.github.io/naturalspeech2/
The code is not yet published, but good news! Soon it will be ready and you will be able to use it from your computer. Currently, the developer lucidrains is replicating the Natural Speech 2 project through the paper published by Microsoft. You can see the latest updates here: https://github.com/lucidrains/naturalspeech2-pytorch
In summary, NaturalSpeech2 is an excellent choice for those seeking a realistic and emotionally expressive voice in their TTS.
Vall-e and vall-ex: Excellent voice quality and open source
While Vall-e and vall-ex are not as well known as NaturalSpeech2, they are excellent options in their own category. Vall-e is a TTS model that generates high quality voices in several languages, including English, German, French, Spanish and more.
One of the biggest advantages of Vall-e is that its code is open source and therefore available to all developers here: https://github.com/lifeiteng/vall-e . This means that anyone with programming experience can modify the model to meet their specific needs.
Vall-ex is an extension of Vall-e that focuses on the expression of emotions in the generated voices. Although not yet released, the model promises to deliver emotionally expressive voice quality similar to NaturalSpeech2.
A series of demos are now available for you to listen to at: https://lifeiteng.github.io/valle/index.html
In summary, Vall-e and vall-ex are excellent choices for those looking for high quality voice quality and the ability to customize their TTS.
Bark: is able to speak for you in multiple languages with your voice simulating your foreign accent.
Bark is a Text-to-Speech (TTS) artificial intelligence that can not only generate realistic multilingual voices, but can also generate other types of sound, such as music, background noise and simple sound effects. Bark has been developed by Suno AI, a company dedicated to the research and development of audio and voice technologies based on artificial intelligence.
What makes Bark especially interesting is its ability to generate music from text. Instead of having to write complex sheet music, users can simply type the text of a song, and Bark takes care of creating the musicality in the corresponding voice.
In addition to music, Bark can also generate background noise and simple sound effects. This is very useful for multimedia content creation, as it allows users to generate custom ambient sounds and sound effects for their videos, podcasts and other types of content.
The source code and a series of demos are available at this link: https://github.com/suno-ai/bark
Another interesting feature of Bark is that it can generate voices in multiple languages with very high sound quality. This is especially useful for those who need to generate content in multiple languages, such as international companies or online content creators who want to reach a global audience.
AudioLM: the artificial intelligence that not only speaks, but also sings and plays personalized sounds.
AudioLM can not only generate high quality multilingual voices, but also includes a GPT chat prompt, where the user can request that a sound with certain characteristics, such as “bird chirps and distant bell echoes” be played, and AudioLM will play it. This is very useful for those who need to generate custom sound effects or ambient sounds for their multimedia projects.
In addition, AudioLM also has the ability to generate different emotions in the voices it creates. As with Vall-e and Vall-ex, the user can choose the emotion he/she wishes to be expressed in the generated voice. These emotions include laughter, crying, sadness, surprise, disgust, drowsiness and many more.
One of the most impressive features of AudioLM is its ability to generate custom voices. Users can train the model with their own voice to create a synthetic version of it. This is very useful for those who want to create multimedia content using their own voice, but do not have the time or resources to record all the narration themselves.
The source code is available here: https://github.com/lucidrains/audiolm-pytorch
In summary, Bark, NaturalSpeech2 and AudioLM are all impressive examples of the speech synthesis technologies that are available today. Each offers unique and advanced features, making them very useful for a variety of multimedia projects.
