Speech-to-text benchmarks
Measured resultsRaspberry Pi 5 · 16 GB RAMAll performance measurements are for this device.
Configuration
Overview
Capability & performance
Benchmark breakdown
Compare models
Word error rate ↓ · processing time ↓Listen
Unedited speech samplesListen for pronunciation, missing words, pauses and voice quality. Transcription scores do not measure how natural a voice sounds.
Test details
Reading the results
Benchmark progress
Researched models
Models we plan to test. Results appear after their benchmarks finish.
Results table · 21 configurations
Overall recognition averages English word error rates from LibriSpeech clean, LibriSpeech other and FLEURS equally. German uses FLEURS alone. A Pi 5 processing factor below 1 means faster than playback. A dash means no matching measurement is available.
| Model | Overall English WER · % ↓ | Overall German WER · % ↓ | Pi 5 processing factor ↓ | Peak memory · MiB ↓ |
|---|---|---|---|---|
| Whisper Tiny | 14 | 33 | 0.191 | 963 |
| Moonshine Tiny | 11 | — | 0.054 | 332 |
| Parakeet TDT 0.6B v3 | 5.4 | 7.9 | 0.112 | 2,271 |
| Whisper Base | 9.6 | 20 | 0.365 | 1,516 |
| Moonshine Base | 8.8 | — | 0.097 | 552 |
| Whisper Small | 6.3 | 11 | 1.255 | 3,767 |
| Underdog Bark 3.0 | 4.9 | — | — | — |
| Fluid Speech Mini | 4.1 | — | 0.33 | 480 |
| Whisper Tiny · AI HAT | 42 | 58 | 0.154 | 558 |
| Whisper Base · AI HAT | 29 | 54 | 0.293 | 608 |
| Gemma 4 E2B IT | 8.6 | 9.6 | 1.591 | 5,654 |
| Gemma 4 E4B IT | 7.7 | 9.9 | 3.184 | 8,231 |
| Qwen3-ASR 0.6B | 4.3 | 8.8 | 0.493 | 2,601 |
| Qwen3-ASR 1.7B | 3.2 | 5.2 | 1.217 | 5,016 |
| Voxtral Mini 3B | 57 | 24 | 15.629 | 6,281 |
| LFM2.5-Audio-1.5B | 4.4 | — | 0.756 | 2,924 |
| Qwen2.5 Omni 3B | 12 | 13 | 9.49 | 6,812 |
| Ultravox 0.5 1B | 251 | 93 | — | 3,476 |
| Qwen2.5 Omni 7B | 58 | 135 | 16.58 | 11,169 |
| LFM3-Audio-2B | 4.9 | 6.6 | 0.715 | 3,100 |
| LFM3-Omni-2B | 4.8 | 7.1 | 0.704 | 3,468 |