AI 热点

Reson8 开源 Rumble-2-Flash 语音模型

SiliconANGLE更新于 11月18日 22:52

Reson8 releases Rumble-2-Flash speech model as a free open-weights download - SiliconANGLE

AI 导读

Reson8 发布开源权重语音识别模型 Rumble-2-Flash,支持 40+ 语言。该 500M 参数模型在多语言测试中平均词错率为 7.09%,在低资源语言上优于 Whisper Large V3。模型采用 Apache 2.0 许可免费开放,API 调用起价 €0.01/小时。

正文

Amsterdam-based speech recognition startup Resonate Labs B.V. today released its Rumble-2-Flash transcription model as a free download with open weights.

Designed for fast transcription of prerecorded audio, the 500 million-parameter model supports more than 40 languages. Reson8 is pitching the release at developers and researchers who need speech recognition built for languages other than English.

According to the company, the leading speech models are trained primarily on English, even though only about one in 20 people speak it as a first language. Specialist terms such as medication names are where accuracy is said to fall furthest once speech moves outside English. Reson8’s models are said to be in production use for meeting notetakers, artificial intelligence receptionists, customer support voice agents, medical transcription and captioning.

“Generic speech models are built for English and make noticeably more mistakes in other languages,” said co-founder and Chief Executive Raoul Ritter, who started the company with Thomas Kluiters and Jarno Verhagen. “Languages like Dutch, Swedish and German are spoken by millions, yet they’re an afterthought.”

Reson8 describes Rumble-2-Flash as one of the smallest models on Hugging Face Inc.’s Open ASR Leaderboard. In the company’s own runs of the leaderboard’s multilingual tests, the model averaged a word error rate of 7.09% across seven languages, including German, French and Dutch. Its average error rate in English is claimed to be about 5%.

On a separate test covering 10 lower-resource languages such as Maltese and Basque, Rumble-2-Flash is said to have posted a mean word error rate of 17.6%. OpenAI Group PBC’s Whisper Large V3 had a mean word error rate of 39.82% on the same test, according to Reson8.

Beyond benchmark sets, Reson8’s models are said to be trained for noisy settings, including phone-line audio. The real world “doesn’t happen in a recording studio,” said Kluiters, the company’s head of AI.

Today’s open-weights release follows the model’s Nov. 3 debut in Reson8’s application programming interface, a week before the company launched its flagship Rumble-2 model. Custom vocabulary support, which lets customers add up to 50,000 terms as text without fine-tuning, remains exclusive to Rumble-2.

Hosted on the company’s own graphics processing units in the Netherlands, Reson8’s service is said to retain no audio or transcripts. The company also claims ISO 27001 certification and compliance with the European Union’s General Data Protection Regulation and the bloc’s AI Act.

Available free under an Apache 2.0 license, Rumble-2-Flash can also be accessed through Reson8’s API at prices starting from €0.01 ($0.001) per hour of audio.

Reson8 raised €5 million ($5.5 million) in pre-seed funding in March. Balderton Capital led the round, with NP-Hard Ventures also participating. Speech is “quickly becoming a default interface across software,” according to James Wise, a partner at Balderton.

Image: Reson8

来源

SiliconANGLE