July 6, 2026

AssemblyAI releases Universal-3.5 Pro Realtime streaming speech-to-text model with 4% WER

AssemblyAI has launched Universal-3.5 Pro Realtime, a new streaming Speech to Text model that achieves a 4.1% word error rate (WER) on AA-WER Streaming, providing significant advancements in real-time transcription capabilities. This model improves upon its predecessor, Universal-3 Pro Realtime, by incorporating conversation context that can be updated after each agent turn without reconnecting, enhancing the model’s ability to produce more contextually relevant outputs. Universal-3.5 Pro Realtime is priced at $0.45 per hour and supports 18 languages, a notable increase from the six languages offered previously, positioning it competitively against other realtime speech-to-text models from providers such as Deepgram, Cartesia, and ElevenLabs.

AssemblyAI: AssemblyAI develops speech-to-text and audio intelligence APIs for developers and businesses. The company recently launched Universal-3.5 Pro Realtime as the successor to its prior streaming model, incorporating turn-by-turn conversation context to deliver more relevant transcriptions. This release expands language support and introduces selectable performance modes while maintaining the same pricing structure.
Universal-3.5 Pro Realtime: Universal-3.5 Pro Realtime is AssemblyAI’s newest streaming speech-to-text model optimized for real-time applications. It succeeds Universal-3 Pro Realtime and enables context updates at the start of a session or after each agent turn without requiring reconnection. The model supports three performance modes and expands coverage to additional languages with mid-sentence code-switching.

Competition: Multiple providers including Deepgram, Cartesia, and ElevenLabs offer competing realtime speech-to-text models that trade off accuracy against latency in streaming scenarios.
Context Awareness: Incorporating dynamic conversation context allows speech models to produce more natural and accurate outputs such as recognizing email addresses from primed phrases rather than spelling them out.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like