Washington Washington. OpenAI on Tuesday introduced a number of new tools that will make it easier for developers to create models based on its artificial intelligence technology, as the ChatGPT maker battles tech giants to stay in the generative AI race. The Microsoft-backed startup said a new real-time tool, which is rolling out immediately for testing, will allow developers to create AI voice applications using a set of instructions.

Previously the process required developers to go through at least three steps: first transcribing the audio, then running a generated-text model to answer the query, and finally using a separate text-to-speech model. A large portion of OpenAI's revenue comes from businesses that use its services to build their own AI applications, making the rollout of advanced capabilities a major selling point.

Competition is also increasing with technology giants including Google-parent Alphabet integrating AI models capable of crunching different types of information such as video, audio and text into their businesses. Reuters reported last month that OpenAI expects its revenue to grow to $11.6 billion next year from an estimated $3.7 billion in 2024. The company is also in the process of raising $6.5 billion in funds, which could value it at $150 billion. As part of Tuesday's rollout, OpenAI introduced a fine-tuning tool for models after training, which will allow developers to improve the responses generated by their models using images and text.

This fine-tuning process may include feedback from humans who give the model examples of good and bad answers based on its answers. OpenAI said that using images to fine-tune models will give them stronger image understanding capabilities, enabling applications such as better visual search and better object detection for autonomous vehicles. The startup also unveiled a tool that would allow small models to learn from larger, established models, as well as “prompt caching” that would halve some development costs by reusing pieces of text previously processed by the AI. Is.