SYLIN

ARK-ASR

by Audio8 · Model · Voice & audio

Open multilingual speech-to-text model in 3B and 0.6B sizes; Apache-2.0 weights, with training and inference code on GitHub

AccessOpen weightsNeeds~7.6 GBSize3B parametersLicenseApache-2.0

Evidence

  1. Aug 2
    3B tops the Hugging Face Open ASR Leaderboard English short-form benchmark, about 5% average word error rate
  2. Aug 2
    Arabic, German, Hindi and Spanish came out okay; Russian failed, Portuguese was half English, Thai failed; unsupported languages get hallucinated output instead of an error

Cost to run

Alternatives

Made from