Voice AI Day is baaaaack! 🎙️ 11 cities, 10 Dec
    Koel Labs

    Koel Labs

    Tech
    STT
    Open Source
    Dataset

    Open phonetic transcription models, IPA benchmarks and datasets for accented and atypical speech.

    Founded 2024Kenmore, United StatesLinkedIn
    Koel Labs banner

    About Koel Labs

    Koel Labs: Open Phonetic Transcription for Accented and Atypical Speech

    Koel Labs is a speech research lab building open models, benchmarks and datasets so speech technology works for accented, non-native and impaired speakers. Its XLSR-English-01 model transcribes English audio into the International Phonetic Alphabet, capturing how words are actually pronounced rather than only which words were said.

    Key Features

    • Phonetic Transcription: XLSR-English-01 turns English audio into sequences of International Phonetic Alphabet symbols from a 79-symbol vocabulary, showing how each word was actually pronounced, with optional timestamps for every phoneme.

    • Built for Diverse Speakers: The model is trained on about 41 hours of native, non-native and impaired speech from corpora such as TIMIT, Buckeye, L2-ARCTIC, EpaDB, Speech Ocean, DoReCo and PSST, and is also tested on accents it never saw in training.

    • Three Model Sizes: The full model (1.18 GB) is joined by XLSR English 01 Mini (338.5 MB) and Nano (7.0 MB) for lighter deployments.

    • Open Phonetic Leaderboard: Koel Labs runs a public IPA transcription leaderboard covering more than 50 models on native, non-native and atypical speech. On it, XLSR English 01 reports a 16.89% average phone error rate and 3.28% feature error rate (self-reported, on Koel Labs' own benchmark).

    • Benchmarks and Datasets: An 80-hour benchmark spanning native, non-native and post-stroke speech, plus standardized versions of public speech datasets published on Hugging Face.

    • Open Source: Training code, processing scripts, models and standardized datasets are published openly; XLSR-English-01 is released under AGPL-3.0 on Hugging Face behind a short access request.

    • Drop-In Usage: The model runs through the Hugging Face transformers automatic-speech-recognition pipeline, including on CPU.

    • Peer-Reviewed Research: An Interspeech 2026 paper reports a 2.3x reduction in weighted phone feature error rate over prior systems, and an IEEE SLT 2026 paper compares 16 speech recognition systems across 8 English accent datasets.

    • Slay Speech App: Koel Labs' iOS and Android app uses phonetic transcription to give English learners word-by-word pronunciation scores, short story-based lessons and conversation practice by text or live voice call.

    Use Cases

    • Pronunciation feedback and language learning products that need phoneme-level scoring rather than word-level transcripts.

    • Researchers and clinicians analysing non-standard or impaired speech, such as speech after a stroke.

    • Linguists studying accents and pronunciation with consistent IPA output.

    • Speech teams measuring how their recognition models perform on accented and atypical speakers.

    Pricing

    • Models, code and datasets: free and open source.

    • Slay Speech: $5.99 per month or $59.99 per year in the US, with a 7-day free trial.

    Getting Started

    Koel Labs gives developers and researchers open phonetic transcription that measures pronunciation across accents and speech conditions, with the benchmarks and data to check how any model performs.

    Alexander Metzger

    Alexander Metzger

    Co-Founder & CEO at Koel Labs

    Connect
    Aruna Srivastava

    Aruna Srivastava

    Co-Founder & CTO at Koel Labs

    Connect
    Ruslan Mukhamedvaleev

    Ruslan Mukhamedvaleev

    Co-Founder & CPO at Koel Labs

    Connect