Real-time speech-to-text API

Real-time speech-to-text is an API that transcribes live audio as it's being spoken, returning partial and final transcripts within milliseconds. It's the right choice for voice agents, live captioning, and meeting copilots, where you need text the moment someone speaks, not after the recording ends.

Gladia's real-time API returns partial transcripts in under 103ms, with sub-300ms end-to-end latency from audio in to transcript out — fast enough for voice agents to respond without dead air and for live captions to stay in sync with the speaker.

Yes. Gladia's real-time API covers 100+ languages, with robust automatic language detection and code-switching support, meaning the model adapts mid-conversation, not just between calls. For example, a caller switching between English and Spanish mid-sentence.

Real-time diarization isn't supported yet. A common workaround is to run the live session through real-time transcription for immediate text, then send the recorded audio through the async API afterward to get a speaker-labeled version for summaries, QA, or record-keeping.

Yes. Gladia is SOC 2 Type 2, HIPAA, GDPR, and ISO 27001 compliant, with data residency options in both the US and EU. This applies across both real-time and async infrastructure.

Gladia's real-time transcription costs 0.68€/hour on the pay-as-you-go Self-serve plan, and new accounts get 50€ in free credits, roughly 73 hours of real-time. Enterprise offers custom pricing with fine-tuned models.

Look at three numbers: partial transcript latency (how fast words appear), final transcript latency (how fast they're confirmed), and word error rate on real conversational audio rather than clean benchmarks. Gladia's real-time API returns partials in under 103ms with sub-300ms end-to-end latency, and scores 9.6% WER on real-life speech in our open-source benchmark. Language coverage and code-switching matter too if your users don't all speak English.

Start with endpointing, the silence that closes an utterance. Voice agents need fast turn-taking, so use 0.05 to 0.1 seconds. Call centers and agent assist work well at 0.2 to 0.4 seconds, responsive without cutting speakers off. Meeting recorders can wait 0.3 to 0.5 seconds so people finish their sentences. Live captions sit around 0.3 seconds, with a 5-second maximum utterance length to keep subtitles readable. Turn on partial transcripts whenever you display text live or need to start processing before the speaker finishes. The docs list recommended parameters by use case.

Use real-time when you need text during the conversation: a voice agent replying to a caller, live captions, or agent-assist prompts mid-call. Use async when the recording is finished and you want the full output, including speaker diarization, summaries, and named entity recognition. Many teams use both, real-time during the call and async afterward for the record.