OpenAI unveils Whisper, an open-source speech-recognition system

OpenAI has unveiled Whisper, an open-source speech-recognition system that transcribes across multiple languages.

According to the announcement, 680,000 hours of multilingual and multitask data collected from the internet were used to train the model. This enables the system to recognise unique accents, background noise and technical jargon, researchers said.

The developers said Whisper delivered good speech recognition results on approximately ten languages.

The company says the model could prove useful to AI researchers studying the reliability, capabilities, limitations and biases of contemporary models.

“Whisper could also prove to be a highly useful solution for developers seeking automatic speech recognition, particularly for English-language speech,” OpenAI said.

Researchers acknowledged that the model has limitations, especially in text prediction. Because the training data included noisy data, Whisper may include words in transcripts that were not actually spoken. The developers suggested this stems from the system’s attempt to predict the next word in the audio and decipher the sound itself.

Whisper does not perform equally well across languages. The system is more prone to errors for speakers whose speech is underrepresented in the training data.

The model’s source code is available on GitHub.

In September, OpenAI allowed editing faces in DALL-E 2. However, developers prohibited uploading images of famous people.

In January, the organisation unveiled a less toxic version of GPT-3, which produces fewer offensive expressions, misinformation, and errors overall.

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!