AssemblyAI Launches Multilingual Real-Time Speech Recognition Across Six Languages

AssemblyAI announced multilingual support for its Universal-Streaming speech-to-text model on November 12, 2025, expanding real-time transcription to six languages: English, Spanish, French, German, Italian, and Portuguese. The release eliminates a critical barrier for companies deploying AI voice agents globally by offering all languages through a single unified model at $0.15 per hour, significantly undercutting competitors like AWS Transcribe ($1.44/hour) and Google Cloud Speech-to-Text ($0.96/hour).

Unified Architecture, Not Language Detection

Unlike competing solutions that route requests through language detection gateways, Universal-Streaming processes all six languages through shared architecture trained simultaneously. This design enables native code-switching within single utterances without additional latency or complexity. The model handles mixed-language phrases like "Je voudrais un coffee, s'il vous plaît" without requiring special handling or preprocessing steps.

Performance Benchmarks on Real-World Audio

AssemblyAI tested the model on diverse audio including call center recordings with background noise, medical consultations with domain terminology, and multi-speaker business meetings. The company reports an average Word Error Rate of 11.77% compared to Deepgram Nova-3's 12.76%, with median latency of 303 milliseconds versus Deepgram's 449 milliseconds. These metrics directly impact user experience in production voice agent deployments where conversation flow depends on sub-400ms response times.

Production-Ready Integration

The model ships with punctuation, capitalization, and intelligent endpointing built in, eliminating post-processing pipelines for downstream LLM integration. Developers can integrate through native support in LiveKit, Vapi, Pipecat, and Daily, or directly via WebSocket API by setting a single parameter. The unified pricing model applies identical rates across all languages, removing variable costs that complicate international expansion planning.

Healthcare and Regulated Use Cases

The consistent accuracy across languages addresses cost concerns in regulated industries where transcription errors compound operational expenses. A 3% higher error rate translates to 5-10% increases in human quality assurance costs at scale, particularly problematic in healthcare documentation and financial services where precision requirements are non-negotiable. AssemblyAI positions the multilingual release as eliminating the traditional trade-off between market reach and product quality.