AssemblyAI has introduced Universal-3 Pro, a new speech language model that it describes as the first production-quality, promptable model of its kind. Unlike traditional speech-to-text systems, Universal-3 Pro allows users to guide transcription behavior through natural language prompts, adapting output for specific domains, terminologies, and formats without retraining or complex post-processing.

Core Innovation: Instruction-Driven Transcription

The model is designed to accept prompts that specify context—such as “medical consultation” or “customer support call”—enabling it to adjust transcription style, capture disfluencies, tag non-speech audio events, and label speakers according to defined roles directly during processing. This approach aims to deliver accurate, domain-aware transcripts from the outset rather than relying on downstream correction layers.

Key Features Highlighted:

· Prompt-Controlled Output: Instruct the model to produce verbatim transcripts, clean reads, or include specific filler words, repetitions, and false starts.

· Keyterm Prompting: Provide up to 1,000 domain-specific terms to improve accuracy, with claimed gains of up to 45% on specialized vocabulary.

· Audio Tagging: Control the inclusion of non-speech events like [laughter], [beep], or [silence] through prompting.

· Native Code-Switching: Supports six languages—English, Spanish, French, German, Portuguese, Italian—with automatic detection and preservation of mixed-language speech.

· Intelligent Language Routing: Can be combined with other AssemblyAI models to cover 99 languages overall, automatically routing audio to the best-fit model.

Performance and Pricing Claims:

AssemblyAI states that Universal-3 Pro achieves the lowest word error rate on real-world data at a cost of $0.21 per hour, which it claims is 35–50% lower than competing solutions. The model is trained exclusively on speech data, which the company says reduces hallucinations and improves reliability for production use cases like medical, legal, and customer service transcription.

Availability:

Universal-3 Pro is available now via AssemblyAI’s API. The company is offering free usage throughout February 2026 (up to 5,000 hours) for users to test the model.