Koel Labs
Koel Labs builds speech technology that accurately understands diverse voices, including accented, non-native, and disordered speech. The company develops state-of-the-art phonetic transcription models using the International Phonetic Alphabet, publishes open datasets and benchmarks, and partners with researchers and clinicians on applied speech research. Their Interspeech 2026 paper demonstrates a 2.3× reduction in weighted phone feature error rate over prior systems.
- Artificial Intelligence
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Voice assistants and smart speakers frequently struggle to understand users with accents, speech differences, or non-standard dialects, leaving millions of people frustrated and unheard by the technology they interact with daily. Most speech systems are trained and measured with a tiny selection of available voices that doesn't represent the world.
Solution
Koel Labs is a research lab that collaborates with universities, researchers, and domain experts to create speech technology that accurately understands diverse voices. The company has built a new state-of-the-art model for phonetic transcription using the International Phonetic Alphabet (IPA) vocabulary, published multiple datasets, and developed internal exploratory tools. Their models are tested on speech aphasia data (PSST) and non-native speech spanning Italian, German (ISLE), and Mandarin (SpeechOcean) accented speech. In their Interspeech 2026 paper, they curate an 80-hour benchmark spanning native, non-native, and post-stroke speech, achieving a 2.3× reduction in weighted phone feature error rate over prior systems with ASR pretraining. Koel Labs openly shares its work through technical reports and blog posts, and invites users to join a waitlist for early access to research previews.
Target Audience
Primary users are researchers, developers, and language learning application builders who need accurate phonetic transcription and inclusive speech recognition for diverse accents and speech patterns. The company also partners with clinicians and domain experts on applied speech research.
Features
- State-of-the-art phonetic transcription model using the International Phonetic Alphabet (IPA) vocabulary to model accent variation
- Validated on diverse speech datasets including aphasia data (PSST), non-native Italian and German (ISLE), and Mandarin accented speech (SpeechOcean)
- Published multiple open datasets and technical reports on phonemic transcription metrics and model architectures
- Technical deep-dives on Wav2Vec2's CNN feature extractor and Transformer architecture, including CTC loss for alignment
- Open-source Hugging Face leaderboards for benchmarking phonemic transcription models
- Early access waitlist for research previews of language tools