Performance

Latency Metrics

Time-to-First-Transcript (TTFT)

Our Pulse STT model provides State of the art TTFT latency of ~64ms, which is one of the least in the world.

Accuracy Metrics

Word Error Rate (WER)

All models were evaluated on the FLEURS dataset, a standardised multilingual speech benchmark ensuring fair cross-model comparison.

Throughput

Requests Per Second

Throughput varies based on audio length, format, and server load

Performance by Audio Format

Linear16 (PCM)

• Latency: Lowest (~64ms)
• Accuracy: Highest
• Bandwidth: Highest
• Best for: High-quality applications

Opus

• Latency: Low (~70-80ms)
• Accuracy: High
• Bandwidth: Low
• Best for: Browser/mobile applications

FLAC

• Latency: Medium (~80-90ms)
• Accuracy: Highest
• Bandwidth: Medium
• Best for: Archival/quality-critical use cases

μ-law

• Latency: Low (~65-75ms)
• Accuracy: Good
• Bandwidth: Lowest
• Best for: Telephony applications

Performance by Language

High-Performance Languages

• Italian: 4.2% WER, ~64ms latency
• English: 5.1% WER, ~64ms latency
• Spanish: 5.4% WER, ~64ms latency
• Portuguese: 7.1% WER, ~64ms latency
• German: 8.5% WER, ~64ms latency
• French: 9.2% WER, ~64ms latency

Regional Variations

• Indian Languages: 10-15% WER, ~90-100ms latency
• Eastern European: 9-12% WER, ~85-95ms latency

Feature Impact on Performance

Diarization

• Latency Impact: +10-20ms
• Accuracy Impact: Minimal
• Use When: Multiple speakers present

Word Timestamps

• Latency Impact: +5-10ms
• Accuracy Impact: None
• Use When: Timing information needed

Emotion Detection

• Latency Impact: +15-25ms
• Accuracy Impact: None
• Use When: Emotion analysis required

Optimization Tips

• Use 16kHz sample rate for optimal balance
• Choose linear16 format for lowest latency
• Enable only needed features to reduce latency
• Batch process when latency isn’t critical

Next Steps

• Metrics Overview.
• Evaluation Walkthrough.
• Best Practices.