The Problem

Pocket records in the real world. Far-field rooms, phone calls, several people talking at once. Generic speech-to-text falls apart exactly there: wrong names, missing words, no sense of who said what.

And the stakes are high. Therapy sessions and financial calls carry sensitive data that has to be redacted, not just transcribed. Users span 10 languages across 16 markets, so one model has to hold accuracy everywhere.

A transcript users can't trust breaks everything downstream: the summary, the action items, the follow-up.

The Solution

Pocket runs Smallest AI's Pulse STT as the transcription layer in its batch pipeline. Because Pocket processes complete files rather than live streams, it uses Pulse in pre-recorded mode, where accuracy is highest.

One pass returns everything: transcript, speaker labels, PII and PCI redaction, punctuation, timestamps, and noise handling. Automatic language detection covers all 10 market languages. Nothing to stitch together as volume grows.

The Results

The following results are;

• 9.63% WER in the real world vs Deepgram's ~28%

• Every transcript ships with speaker labels, redaction, punctuation, and timestamps in a single pass
  
  Pulse holds accuracy across all of Pocket's market languages:
  
  insert table on multilingual benchmarks here
  
  And it stays ahead as conditions degrade. WER by noise band:
  
  insert table on noise-band benchmarks here
  
  Batch full-file processing scales with volume with no per-stream limits, so quality stays consistent across every market and use case.
  
  Building something that depends on accurate transcription? See the Pulse model card or talk to our team.