読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
トピックDisconfirmAI 編集部更新日 2026-11-21 18:44

Reson8 Rumble-2-Flash

3 ステップ1 件4 ソース

Amsterdam-based speech AI startup Reson8 (Resonate Labs B.V.) has released Rumble-2-Flash, a 500-million-parameter open-weight automatic speech recognition (ASR) model optimized for non-English languages. The company argues that leading models are trained primarily on English, resulting in higher error rates for other languages, particularly with specialized terms like medication names. Rumble-2-Flash is designed for fast transcription of prerecorded audio and supports over 40 languages.

According to Reson8's internal tests on the Hugging Face Open ASR Leaderboard, the model achieved an average word error rate (WER) of 7.09% across seven languages including German, French, and Dutch, and approximately 5% in English. On a separate test covering ten lower-resource languages such as Maltese and Basque, Reson8 claims a mean WER of 17.6%, compared to 39.82% for OpenAI's Whisper Large V3. The model is available under an Apache 2.0 license and can be accessed via Reson8's API starting at €0.01 per hour of audio.

The release follows the model's debut in Reson8's API on November 3, 2026, and precedes the launch of their flagship Rumble-2 model. While Rumble-2-Flash focuses on speed and cost-efficiency for general transcription, it lacks features exclusive to the larger Rumble-2 model, such as custom vocabulary support for up to 50,000 terms without fine-tuning. Reson8 states its infrastructure is hosted on its own GPUs in the Netherlands, claiming zero data retention and compliance with GDPR and the EU AI Act.

経緯

  1. 2026-11-03

    November 3, 2026: Rumble-2-Flash debuts in Reson8's application programming interface (API).

    reson8.dev ↗
  2. 2026-11-10

    November 10, 2026: Reson8 launches its flagship Rumble-2 model.

    siliconangle.com ↗
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク
  • 2026-11-18

    November 18, 2026: Reson8 releases Rumble-2-Flash weights as a free download under the Apache 2.0 license.

    huggingface.co ↗
  • 未解決の問題

    • How do independent third-party benchmarks compare to Reson8's self-reported performance metrics?
    • What is the actual latency and accuracy trade-off when running the model on consumer-grade hardware versus cloud GPUs?
    • Will Reson8 introduce speaker diarization or real-time streaming capabilities to the open-weight Flash model in future updates?

    ソース

    • siliconangle.comSiliconANGLE↗
    • huggingface.coHugging Face Model Card↗
    • huggingface.coHugging Face Blog↗
    • reson8.devReson8 Blog↗

    このトピックでのカバー範囲

    • Dutch Speech AI Startup Reson8 Releases Open-Weight Multilingual Model Rumble-2-Flash2026-11-19ニュース

      Reson8 releases Rumble-2-Flash, a 500-million-parameter open-weight ASR model optimized for non-English languages.