Connecting the Dots: Enduring Performance Gaps in ASR Systems Between Standard and Minority English Varieties
Accepted Paper to IEEE SLT 2026
We evaluate 16 speech recognition systems released between 2021 and 2025 across 8 datasets of English accent varieties. The gap between Mainstream US English and minority accents is not reliably closing over time, and minority accents can regress sharply between model generations, suggesting standard benchmarks implicitly favor mainstream varieties.
Scaling Human and G2P Supervision for Robust Phonetic Transcription
Accepted Paper to Interspeech 2026
We are presenting at Interspeech 2026 about how phonetic transcription systems scale with expert human labels versus G2P-generated supervision. Across native, non-native, and post-stroke speech, we find that G2P helps only when expert annotation is scarce, while diverse human labels and ASR pretraining drive stronger robustness and reduce weighted phone feature error by 2.3x.
Speak Like The Stars: Pronunciation Mastery Through Movies.
Interspeech Conference Science Fair
We presented our work on making speech pathology more accessible, insights and statistics into our latest phonetic model, and a demo of our work-in-progress pronunciation platform.
Speech Technology Built for Everyone, Everywhere.
University of Washington Research Symposium
We presented our work on making speech pathology more accessible, insights and statistics into our latest phonetic model, and a demo of our work-in-progress pronunciation platform.
Leaderboard
Setting a new standard in phonetic transcription
| Koel Labs XLSR English 01 | 20.59% / 3.31% |
| Koel Labs XLSR English 01 Mini | 20.80% / 3.40% |
| Qwen3.8-Omni-Flash | 38.79% / 5.97% |
| Facebook Wav2Vec2 53 Espeak | 114.40% / 7.68% |
| Facebook Wav2Vec2 60 Espeak | 116.98% / 7.96% |
| MrRubino XLSR 53 L2-Arctic | 58.83% / 8.12% |
| Koel Labs XLSR English 01 Nano | 37.63% / 8.44% |
| PhoneticXeus | 49.98% / 8.46% |
| Vitouphy XLSR TIMIT | 59.91% / 8.55% |
| WavLM HuPER CTC | 39.12% / 8.55% |
| ZIPA Large | 45.82% / 8.80% |
| ESPnet POWSM | 50.41% / 8.88% |
| Ginic XLSR 53 Buckeye | 42.12% / 8.89% |
| Facebook XLSR 53 Phon | 50.01% / 9.35% |
| DahmL XLSR MFA | 122.82% / 10.05% |
| Wav2Vec2-base TIMIT arpabet | 45.14% / 10.09% |
| HuBERT Phoneme | 42.14% / 10.20% |
| Koel Labs B0 | 44.36% / 10.30% |
| Facebook XLSR 53 BABEL | 54.83% / 10.89% |
| MultiBridge EEA Wav2Vec2-base | 44.72% / 11.70% |
| Speech31 XLSR English IPA | 117.12% / 12.37% |
| Speech31 Wav2Vec2 TIMIT-IPA3 | 68.59% / 12.64% |
| Chihiro Taguchi XLSR japlmthufielta | 73.07% / 12.89% |
| Speech31 WavLM English IPA | 118.56% / 13.29% |
| Snu Nia Wav2Vec2 TIMIT | 141.47% / 13.46% |
| Speech31 TIMIT-IPA2 | 141.47% / 13.46% |
| Speech31 Hubert English IPA | 109.03% / 14.16% |
| Allosaurus | 70.21% / 14.21% |
| SNU-SLP Lab XLSR 53 TIMIT | 71.75% / 14.71% |
| Jubliano XLSR International 1.5 | 83.61% / 16.46% |
| WhisperPPT | 68.12% / 17.01% |
| Vikramanantha Wav2Vec2-base LJSpeech Gruut | 141.09% / 22.98% |
| Seanghay Wav2Vec2-base Khmer | 115.72% / 25.15% |
| Cole Robertson Wav2Vec2 Ogma | 116.51% / 28.03% |
| Aditya Wav2Vec2 Librispeech | 164.09% / 55.05% |
| Excalibur12 Wav2Vec2 TIMIT Simplified | 283.05% / 57.73% |
| Excalibur12 Wav2Vec2 lv60 TIMIT | 289.38% / 59.85% |
Last updated: September 20, 2026 from https://huggingface.co/spaces/KoelLabs/IPA-Transcription-EN
Research and Collaboration
Our Way Forward
Driven by a commitment to collaborative science, we advance open-source research in language technologies while building practical, engaging solutions. This dedication stems from a core belief: transformative innovation builds upon collective knowledge.
Through open-source contributions, our organization accelerates progress in linguistic technology, shaping a future where language learning knows no bounds.
We work with linguists, natural-language processing researchers, and research labs worldwide. If you're interested in our work, please contact us to explore collaboration opportunities!
