AssemblyAI-universal-3-pro
Neural Network
Specifications
- Input
- audio
- Output
- text
Prompt cost
(per 1M tokens)
How it works AssemblyAI-universal-3-pro?
Frequently asked questions about AssemblyAI-universal-3-pro
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is a speech recognition model: you upload audio, and you get text as output. It is suitable for transcribing interviews, calls, podcasts, lectures, and voice messages. The BIG_AUDIO_TO_TEXT mode is designed for long recordings, so large files can be submitted in their entirety without manual cutting.
According to our data, the output limit is 4096 tokens, and the context is 4095. For transcription, this means that a very long recording will be returned in parts: process the audio in fragments and stitch the result together. The volume of the final text depends on how much speech is actually contained in your file.
A separate reasoning mode is not noted in our data, and it is not needed for the task: the model converts speech to text, it does not reflect on the answer. If you need conclusions, a summary, or an analysis of the transcription, pass the finished text to a text model — in BotHub, they are available in the same window.
Function calling and structured output are not noted in our data — this is a transcription model, not a tool for agent scenarios. However, you can access it via the unified OpenAI-compatible BotHub API, and structure the transcription according to a schema using another model in the same integration.
Audio — yes, this is the model's main input, including long recordings. It does not accept images or video: there is no image generation or frame analysis here. If you need to transcribe a video, extract the audio track from it and send that instead.
Universal-3-pro from AssemblyAI is a multilingual speech recognition model; there are no specific metrics for individual languages in our data, so it is easier to check the quality with your own recording. Access to it in BotHub works from Russia without a VPN or foreign card, with payment in rubles upon usage.