Speech recognition engines
Choose a recognition engine for each spoken language. Incoming speech can use the same local engines to transcribe other people.
If you are unsure
| Spoken language | Recommendation |
|---|---|
| Japanese | Parakeet Japanese TDT or Haas-ja |
| English | Parakeet English |
| Other languages or automatic language detection | Qwen3-ASR or Whisper turbo |
| Lower-spec PCs | Web Speech (cloud) |
Engine overview
| Engine | Runs on | Languages | Download | Time relative to speech length | Features |
|---|---|---|---|---|---|
| Parakeet Japanese TDT | This PC | Japanese | ~1.0GB | ~3.5% | Fast and accurate; punctuation included |
| Haas-ja | This PC | Japanese | ~480MB | ~13% | Slightly more accurate than Parakeet Japanese TDT on normal conversation; generates kanji and punctuation itself |
| Parakeet English | This PC | English and other European languages | ~670MB | ~5% | Fast; punctuation and capitalization. No Japanese, Chinese or Korean |
| Qwen3-ASR | This PC | 52 languages including Japanese, English, Chinese and Korean | ~1.0GB | ~10–20% | Many languages in one model, with automatic language detection |
| Whisper turbo | This PC | Japanese, English, Chinese, Korean and more | ~1.0GB | ~30% | Automatic language detection; slower than the other local engines |
| Web Speech (Microsoft) | Cloud | Many languages | None | — | No model download; needs internet and sends audio to Microsoft |
| Google (unofficial, experimental) | Cloud | Many languages | None | — | Not an official API; may stop working without notice. Audio goes to Google |
- Local engines use only CPU, not GPU. Audio stays on your PC.
- Times are indicative measurements on the developer’s Ryzen 7 7700. “~5%” means about 0.5 seconds to recognize 10 seconds of speech.
- Haas-ja’s ~13% is rounded from the previous relative measurement (~3.6× Parakeet Japanese TDT) using ~3.5% as the baseline.
- Only Qwen3-ASR and Whisper turbo allow the language to be set to Auto.
Results on synthesized speech
Each engine transcribed the same Qwen3-TTS/Irodori-TTS audio, and results were compared to the source text. This helps show differences in engine behavior on identical audio.
Cells show “mean error rate · exact matches/30.” Lower error rates are better.
Japanese (pronunciation-based comparison)
The app corrects Parakeet Japanese TDT’s kanji afterward, so both reference and recognized text are converted to hiragana here.
| Audio source | Whisper turbo | Qwen3-ASR | Parakeet Japanese TDT | Haas-ja |
|---|---|---|---|---|
| Irodori-TTS | 5.6% · 25/30 | 13.8% · 21/30 | 6.9% · 20/30 | 4.5% · 25/30 |
| Qwen3-TTS | 14.9% · 23/30 | 17.3% · 19/30 | 11.1% · 23/30 | 9.7% · 23/30 |
Other languages (Qwen3-TTS audio)
| Language | Whisper turbo | Qwen3-ASR | Parakeet English |
|---|---|---|---|
| English | 2.1% · 27/30 | 0.8% · 28/30 | 2.9% · 25/30 |
| Chinese | 5.4% · 25/30 | 0.3% · 29/30 | Not supported |
| Korean | 2.4% · 27/30 | 4.0% · 24/30 | Not supported |
| German | 3.9% · 26/30 | 4.4% · 25/30 | 4.7% · 23/30 |
| Spanish | 4.2% · 27/30 | 2.8% · 28/30 | 3.9% · 25/30 |
| French | 6.6% · 24/30 | 6.1% · 27/30 | 2.5% · 26/30 |
| Italian | 6.4% · 24/30 | 11.2% · 21/30 | 5.9% · 21/30 |
| Portuguese | 6.2% · 26/30 | 7.4% · 24/30 | 6.5% · 21/30 |
| Russian | 2.6% · 26/30 | 9.3% · 22/30 | 4.4% · 25/30 |
Limitations
- These are synthetic voices, not human microphone recordings. Noise, real voices and individual speaking habits can change results.
- Each language uses 30 short everyday sentences, AI-authored without native-speaker review.
- Japanese, Chinese and Korean errors are counted per character; other languages per word. Number formatting (“3” versus “三”) and paraphrases also count as errors.
- App post-processing such as kanji conversion and punctuation correction was not applied.
- Cloud engines (Web Speech and Google) were not tested.