AI is not just limited to computation and analysis. They are very wide and can be used in a wide variety of areas. The American company Microsoft, for example, is working on an unusual project, the first results of which can be heard right now. Yes, to hear, because we are talking about artificial, but already realistic vocals.
The project was named DeepSinger and is the first system in the world that learns to sing using songs posted on the Web. The algorithm of its work is complicated, but it can be described in a few phrases in a simplified form. First, the song is decomposed into instrumental accompaniment and vocals. The vocals are then broken down into sentences. Phonemes are distinguished from them. Next, another proprietary Microsoft FastSpeech technology comes into play, which forms the “singing model”.
According to the developers, it is too early to talk about the readiness of DeepSinger for commercial use. Work towards improving algorithms for processing audio data will continue. However, it is already possible to evaluate the capabilities of the new system by listening to several vocal fragments created by it in Chinese and English. In the future, there will be work to be done on the quality and realism of artificial singing.
Share what you read – it’s not difficult, but very important for us! Thank you.