Revolutionary Speech‑to‑Text: AssemblyAI’s New STT‑3.7 Models Power Accurate, Multilingual Transcriptions

Introducing AssemblyAI STT‑3.7

AssemblyAI has unveiled the STT‑3.7 family of speech‑to‑text models, offering a suite of highly accurate transcription solutions for recorded audio, live conversations, and voice‑agent interactions. Built on advanced neural architectures, these models deliver industry‑leading accuracy across a broad range of languages and real‑world scenarios.

Three Tailored Variants

  • AssemblyAI‑STT‑3.7 (Recorded Audio) – The flagship model for pre‑recorded content. It excels at recognizing entities, rare words, and domain‑specific terminology, thanks to an updated speaker diarization engine that accurately attributes speech to participants, handles code‑switching, and captures names and numbers with high fidelity.
  • AssemblyAI‑STT‑3.7‑Realtime – Optimized for live streams such as meetings, conference calls, and live captions. The model outputs real‑time speaker labels that are finalized at the end of each speaking turn, and it automatically detects and supports 32 native languages within a single stream.
  • AssemblyAI‑STT‑3.7‑Realtime‑Agent – Designed for voice‑assistant contexts. It prioritizes short, actionable replies and reduces background speech, ensuring that agent‑triggered actions receive clean, precise transcriptions. Features like Voice Focus enhance performance in noisy environments.

Key Innovations Across the Family

  • Unmatched Accuracy – Benchmark tests show STT‑3.7 surpasses previous models in diverse datasets, delivering higher word‑error rates across all supported languages.
  • Multilingual Reach – Every variant natively supports 32 languages with seamless code‑switching. The recorded version offers a 99‑language fallback for broader coverage.
  • Enhanced Speaker Diarization – Fewer errors, faster processing, and precise speaker counts improve the clarity of multi‑speaker transcripts.
  • Contextual Prompting – A 400‑character prompt guides the model toward specific vocabulary, while a 1,000‑word keyterm boost sharpens recognition of jargon and custom terms.
  • Voice Focus – Exclusive to the Realtime‑Agent model, this feature filters background noise, sharpening user intent.
  • Medical Mode – An add‑on that specializes in healthcare terminology, boosting accuracy for medication names and clinical dialogue across all STT‑3.7 models.

Developer Integration and Pricing

The STT‑3.7 models are accessible through AssemblyAI’s RESTful API, allowing developers to embed robust voice AI into products ranging from customer support tools to accessibility services. Pricing tiers for real‑time use are set at $0.30 per hour for the Realtime model and $0.45 per hour for the Realtime‑Agent model, reflecting the advanced processing required for live speech.

Why Choose STT‑3.7?

With its combination of high accuracy, multilingual capability, and specialized features like Voice Focus and Medical Mode, the STT‑3.7 family represents a significant leap forward in speech AI. It enables businesses and developers to deliver reliable, real‑time transcription services without compromising on clarity or contextual understanding.

AssemblyAI’s transition from Universal‑3.5 Pro and Universal‑3.6 Pro Realtime to the STT‑3.7 line marks a new era for scalable, high‑performance speech processing.

Leave a Reply

Your email address will not be published. Required fields are marked *

Close filters
Products Search