Future Developments<\/h2>\n
\u201cWe\u2019re going to see more and more agents emerge in the coming months,\u201d Godement told TechCrunch during a briefing. \u201cAnd so, the general theme is about helping customers and developers leverage agents that are useful, available, and accurate.\u201d<\/p>\n
Text-to-Speech Model<\/h2>\n
OpenAI claims its new text-to-speech model, “gpt-4o-mini-tts,” delivers more nuanced and realistic-sounding speech and is more “steerable” than its previous speech-synthesizing models. Developers can instruct gpt-4o-mini-tts using natural language, such as “speak like a mad scientist” or “use a serene voice, like a mindfulness teacher.” <\/p>\n
Voice Samples<\/h2>\n
Here\u2019s an example of a “true crime-style,” weathered voice: <\/p>\n
And here\u2019s a sample of a female “professional” voice: <\/p>\n
Developer Control<\/h2>\n
Jeff Harris, a member of the product staff at OpenAI, told TechCrunch that the goal is to let developers tailor both the voice “experience” and “context.” <\/p>\n
\u201cIn different contexts, you don\u2019t just want a flat, monotonous voice,\u201d Harris said. \u201cIf you\u2019re in a customer support experience and you want the voice to be apologetic because it\u2019s made a mistake, you can actually have the voice convey that emotion \u2026 Our big belief here is that developers and users want to really control not just what is spoken, but how things are spoken.\u201d<\/p>\n
Speech-to-Text Models<\/h2>\n
OpenAI’s new speech-to-text models, “gpt-4o-transcribe” and “gpt-4o-mini-transcribe,” effectively replace the company’s older Whisper transcription model. Trained on diverse, high-quality audio datasets, these models can better capture accented and varied speech, even in chaotic environments, according to OpenAI.<\/p>\n
Improved Accuracy<\/h2>\n
They\u2019re also less likely to hallucinate, Harris added. Whisper was known to fabricate words and passages<\/a> in conversations, introducing irrelevant information into transcripts.<\/p>\n \u201c[T]hese models are much improved versus Whisper on that front,\u201d Harris said. \u201cMaking sure the models are accurate is completely essential to getting a reliable voice experience, and accurate means that the models are hearing the words precisely and aren\u2019t filling in details they didn\u2019t hear.\u201d<\/p>\n However, the performance may vary depending on the language being transcribed. <\/p>\n According to OpenAI\u2019s internal benchmarks, gpt-4o-transcribe, the more accurate of the two transcription models, has a word error rate approaching 30% for Indic and Dravidian languages such as Tamil, Telugu, Malayalam, and Kannada. This means three out of every 10 words from the model will differ from a human transcription in those languages.<\/p>\n In a departure from its usual practice, OpenAI doesn\u2019t plan to make its new transcription models openly available. The company had previously released new versions of Whisper<\/a> for commercial use under an MIT license.<\/p>\n Harris stated that gpt-4o-transcribe and gpt-4o-mini-transcribe are \u201cmuch bigger than Whisper\u201d and thus not suitable for an open release.<\/p>\n \u201c[T]hey\u2019re not the kind of model that you can just run locally on your laptop, like Whisper,\u201d he continued. \u201c[W]e want to make sure that if we\u2019re releasing things in open source, we\u2019re doing it thoughtfully, and we have a model that\u2019s really honed for that specific need. And we think that end-user devices are one of the most interesting cases for open-source models.\u201d<\/p>\nLanguage-Specific Performance<\/h2>\n
![\"OpenAI](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2025\/03\/Screenshot-2025-03-20-at-10.38.42AM.png?w=680\")
Availability of New Transcription Models<\/h2>\n
Update<\/h2>\n