Pulse — Speech-to-Text by Smallest AI

Overview

Pulse provides streaming and pre-recorded speech-to-text for voice agents, live captions, meetings, and recorded conversations. Its real-time API uses WebSocket connections; recorded files use an HTTP endpoint. Smallest AI reports sub-100ms time to first token at one concurrent request. Latency varies with concurrency and workload.

Key features

Transcription capabilities include speaker labels, word and utterance timestamps, keyword boosting, language detection, and inverse text normalization. Language availability differs between streaming and batch modes and between deployment regions; consult the current Pulse model card for the supported language matrix. Pulse Pro is a separate English-only model for pre-recorded audio. It is designed for workloads such as call archives, meetings, and financial-audio transcription. Its model card reports 5.42% average word error rate on the ESB benchmark. Benchmark results depend on the dataset and evaluation setup; Pulse Pro currently does not support the live streaming endpoint. PII and PCI redaction are optional request controls for masking personal and payment-card information in supported transcripts. The documented redaction languages are English and Hindi. Teams handling regulated data should confirm the applicable deployment, contractual terms, and security documentation with Smallest AI rather than infer compliance from a feature label.

Use cases

Smallest AI publishes security information through its Trust Center and offers on-premise deployment for enterprise customers. These options should be assessed against the requirements of the intended application.

Related links

• Pulse model card
• Pulse Pro model card
• Streaming API capabilities
• Smallest AI Trust Center
• On-premise deployment