Smallest AI's Pulse delivers managed real-time speech-to-text at ~64ms time-to-first-token and 5.42% average WER (#2 on the Open ASR Leaderboard), versus ~7.4% for open-source Whisper large-v3. Self-hosting Whisper removes license fees but adds streaming, accuracy, and GPU-ops burden; Pulse trades a per-minute fee for managed low-latency performance.
Pulse Speech-to-Text
Real-time transcription with industry-leading accuracy and ~64ms time-to-first-token, built for live voice agents.
30+ Languages
Accurate transcription across 30+ languages, with native English performance ranked #2 on the Open ASR Leaderboard.
Enterprise-Grade Compliance
SOC 2 Type II, GDPR, ISO 27001, and HIPAA-ready — with a Business Associate Agreement available for healthcare deployments.
Features
Benchmarks
WER % on the Hugging Face ESB benchmark (9 English datasets, streaming). Lower is better. Pulse is competitive on aggregate and leads on meeting/noisy audio; AssemblyAI leads on clean read speech. Source: HF ESB, internal evaluation.
Frequently
asked questions
Is open-source speech recognition free?
Does Whisper support real-time streaming?
When does self-hosting open-source STT make sense?
How accurate is Pulse versus open-source models?
Can Pulse run on-premise like a self-hosted model?
What does moving from self-hosted Whisper to Pulse involve?