OpenAI has released upgraded transcription and voice-generating AI models, boasting significant improvements over previous versions. Aligning with its vision of more autonomous systems, OpenAI aims to create models that can independently perform tasks for users. Olivier Godement, OpenAI’s Head of Product, describes these models as a step towards building chatbots capable of engaging in meaningful conversations with customers.
The new text-to-speech model, called “gpt-4o-mini-tts,” is designed to produce more realistic and nuanced speech while allowing developers to customize tones and styles. Users can guide the model on how to deliver lines, whether adopting a quirky style reminiscent of a “mad scientist” or a calm demeanor like a mindfulness teacher. Jeff Harris, a member of OpenAI’s product team, emphasizes the goal of enabling developers to craft both voice experience and emotional context for better user interactions.
In addition to the text-to-speech upgrades, OpenAI has introduced “gpt-4o-transcribe” and “gpt-4o-mini-transcribe” as replacements for the outdated Whisper transcription model. These new models were developed using diverse and high-quality audio datasets to improve the accuracy of speech recognition, especially in challenging environments. Harris notes a reduction in errors, claiming these models will not fabricate information as Whisper often did.
Despite these advancements, challenges remain, particularly for Indic and Dravidian languages, where OpenAI reports a word error rate nearing 30%. Interestingly, the new transcription models will not be available for open-source usage, diverging from OpenAI’s tradition of releasing models under open licenses. Harris explains that these larger models are not suitable for local machine operation, prompting a more considered approach to their release.
