Pulse

WER on the English subset of FLEURS across providers in streaming mode. Lower is better.

A note on audio amplitude normalization

Audio amplitude normalization materially changes WER on FLEURS. Most competitors benchmark on raw FLEURS - which has variable, often low amplitude - without normalizing peak audio to −10 dBFS. This makes some models look much better than they actually are. Pulse is stable across all amplitude regimes.

WER on FLEURS in streaming mode, broken down by language family. Lower is better.

WER on FLEURS in pre-recorded mode (full-file upload). Lower is better.

WER across seven Hindi datasets covering read speech, conversational speech, telephony / contact-center audio, and noise-augmented variants. Compared against IndicWhisper, Sarvam Saaras v3, scribe v2, and Deepgram Nova-3. Lower is better.

For the full breakdown including training-data and evaluation-protocol notes, see the Performance page.

A Hugging Face benchmark suite aggregating 9 English speech datasets across diverse domains (audiobooks, parliament, meetings, finance, etc.) to test STT generalization. Lower WER is better.

Evaluated on the open-source Hugging Face ESB datasets. Numbers from internal evaluation.

An open-source robustness benchmark designed to stress-test STT under real-world degraded conditions: clipping, far-field capture, background noise, phone codec compression, reverberation, and accented speech. Lower WER is better. n/a = not supported by that provider.

Evaluated on the open-source WildASR dataset. Numbers from internal evaluation.

WER and Character Error Rate (CER) on live, dual-channel contact center call recordings (agent and customer channels scored separately), covering English and Hindi. Diarization was applied first to cleanly separate agent and customer speech before scoring. Lower is better.

Evaluated on real-world telephony call audio. Numbers from internal evaluation.

Pulse leads across both languages on real production call audio, extending its edge on Hindi CER in particular. For the full breakdown, see the Performance page.

WER for the four East Asian languages newly enabled on the streaming endpoint (us-west-2). Three datasets per language covering read speech (FLEURS), conversational/crowdsourced speech (Common Voice 25), and language-specific corpora (JSUT, Zeroth-Korean, MDCC, AISHELL-1). Compared head-to-head against Deepgram Nova-3. Lower WER is better.

Pulse averages 10.91% WER vs Deepgram Nova-3’s 15.50% across the four East Asian languages - Pulse leads on 10 of 12 dataset rows, with the largest gains on Japanese CV-25 and Cantonese CV-25.

These four languages stream from wss://api.us.smallest.ai/waves/v1/stt/live?model=pulse only (US region). See the streaming Asian-language documentation for the region-routing details.

Not a public dataset. The English audio is sliced by perturbation type (Noise, Silence, Telephony 911, Boundary, Disfluency, Long Audios, Repetition, Entity, Accent, Emotion, Speaker Diversity, Speed, Pitch, Volume, Audio Quality) to isolate model weaknesses. Lower WER is better.

Not a public dataset. Hindi audio sliced by perturbation type to isolate model weaknesses. Lower WER is better except for Entity EDR where higher is better (↑).

Pulse ships speaker diarization on both batch and streaming through the same API — no separate diarization vendor to stitch in. See the Diarization page for the response shape and enable flag.

Metric. DER (Diarization Error Rate) = percent of speech time handled wrong (wrong speaker + missed speech + falsely detected speech). Scored with no collar, overlapping speech counted, permutation-optimal speaker matching, duration-pooled via pyannote.metrics. Lower is better.

Streaming DER vs other STT APIs

Per-dataset DER. Lower is better. - means we have not scored ElevenLabs on that dataset yet.

Extended benchmark against the specialist API pyannoteAI (diarization-only, not an STT API): pyannoteAI wins on some meeting-room datasets (ami_ihm 12.8 vs 18.5, voxconverse 7.7 vs 20.8) where speaker-count is high. Pulse remains competitive and wins on French emergency dispatch (simsamu 13.4 vs 14.7).

Streaming DER on Indic languages

Streaming diarization measured end-to-end through the WebSocket API at concurrency 1. Lower is better. Pulse macro-average across all 22 Indic language codes is 23.2 DER; Pyannote Live-1 is 37.1; Deepgram Live is 77.6 (Deepgram Live does not have a first-class Indic pipeline).

Speed (real-time factor)

RTF = processing time / audio length. Lower is faster. 1/RTF is “times faster than real-time.” Measured through each vendor’s API at concurrency 1 with diarization enabled — the round-trip a customer actually sees, not local compute.

Streaming diarization emits speaker labels on is_final: true transcription events with roughly the same latency as transcription itself (~1 s tail from utterance end to final event on typical connectivity).

[*] East Asian languages (zh, yue, ja, ko, multi-asian) are served from the US region only. Connect to wss://api.us.smallest.ai/waves/v1/stt/live instead of wss://api.smallest.ai/... for these.

[**] South Indian languages (ta, te, kn, ml, multi-south-indic) and hi-dev are served from the India region only (wss://api.smallest.ai/...). Requests to wss://api.us.smallest.ai/... are rejected with error code LANGUAGE_NOT_ENABLED_IN_REGION. Contact support to request access.

Pulse runs two independent inference paths. Streaming uses a WebSocket connection and emits partial transcripts in real time (suited for live call transcription, voice assistants, and conversational AI where first-token latency matters). Batch accepts a full audio file over HTTP and returns the complete transcript once processing is done (suited for call recordings, media archives, and any workload where you have the full audio upfront). Both modes support keyword boosting, diarization, word-level timestamps, and PII/PCI redaction. Streaming covers 21 languages; batch covers 12 languages, all of which are also on streaming.

Use a specific language code (e.g. hi, es, en) whenever you know the language of the audio. Pulse optimises directly for that language and handles code-switching with English automatically, so hi covers Hindi/English mixed speech without needing an aggregator. Use north_indic or multi-asian only when the source language is genuinely unknown or mixed across several languages. Auto-detection adds a small accuracy overhead compared to an explicit code.

The complete request/response schema, all query parameters, error codes, and WebSocket protocol details are in the API Reference. For step-by-step setup, see the Realtime quickstart for streaming and the Pre-recorded quickstart for batch.

Pulse supports both streaming and batch transcription across a broad multilingual set with regional aggregators. Pulse Pro is English-only and batch-only, but achieves higher accuracy: tied for #2 on the Open ASR Leaderboard at 5.42% average WER vs Pulse’s 5.55% on English FLEURS. Use Pulse for live streaming, multilingual audio, or latency-sensitive workloads. Use Pulse Pro for high-volume English batch transcription where accuracy is the top requirement.

Pulse applies built-in redaction on both streaming and batch surfaces (no preprocessing required). PII redaction masks personal identifiers such as names, phone numbers, email addresses, and SSNs. PCI redaction masks payment card data including card numbers, CVVs, and expiry dates. Both are enabled via query parameters in the API request. Redacted content is replaced with a placeholder in the transcript. The original audio is not retained post-processing. Supported on language=en and language=hi only.