Pulse STT

Pulse STT has been updated with an improved English model. The Pulse Performance page (/models/documentation/speech-to-text-pulse/benchmarks/performance) has been refreshed with new WER for the standard public benchmarks - FLEURS, HuggingFace ESB, and WildASR - and improves on every one of them. WildASR sees the largest single-benchmark move, driven by big gains on the harder degradation subsets (Far-field, Clipping, Reverberation, Noise Gap).

Beyond the public benchmarks, two production-facing failure modes see substantial gains:

• Repeated-entity dictation - spoken digit sequences, alphanumeric IDs, PAN / credit-card / reference-number utterances, and any content with long runs of the same token. WER on this class of audio drops by 54% relatively vs. the previous model. This is the same failure mode where earlier releases would collapse "3 3 3 3 3 3 3" down to a shorter run; the new model preserves run length far more consistently.

• Background-speaker robustness - audio where a secondary speaker or crosstalk is audible under the primary speaker. WER on this class also drops by 61% relatively, so multi-speaker call recordings and open-office capture behave much closer to clean single-speaker audio than before.

These improvements are live behind the existing Pulse STT streaming endpoints - no API changes, no request-side flags to flip.