Eigen RadarAI
Analysis

Reson8 opens the weights of its multilingual speech model

Reson8 released the weights of Rumble-2-Flash, a model that turns prerecorded speech into text in more than 40 languages. The release gives developers another way to deploy a model that the Dutch startup already offers through its own service. The system can detect the language and let developers choose how some spoken words appear in the transcript.

Artificial Intelligence··Night
Low-angle wide view of a Dutch speech-AI workspace with desktop audio playback speakers, an LED waveform deck and a glass board aligning a physical soundwave strip with two rows of blank colour-coded segmentation blocks.

The downloadable model handles recorded speech

Reson8, a Dutch speech recognition startup, released the weights of its Rumble-2-Flash model for developers to run. The weights are the numerical settings the system learned during training. Its focus is prerecorded audio in more than 40 languages. Reson8 had already been offering the model through its own service. Developers now have a separate way to use it.[1], [2]

Developers can set the transcript style

The recognizer can work out which language people are speaking or accept a language chosen in advance. Developers can also choose whether it puts numbers into the transcript as digits or words, keeps filler words and writes down repeated words as spoken. The example published on Hugging Face, a platform for sharing AI models, shows that the system can take audio from a file and handle more than one sample at a time. It can also accept audio from a link.[1]

A larger model helped train the smaller one

Reson8 trained the small model with help from its larger Rumble-2. The smaller system rendered its own transcript first. The larger one then supplied information about what would best fit at each point in the transcript. The training included mistakes the smaller system actually made. The two models break text into different pieces. The team used a method to align those pieces during training.[1]

References

  1. News sourceHugging FaceReson8 releases Rumble-2-Flash weights under Apache 2.0↩1↩2↩3
  2. News sourceSiliconANGLEReson8 releases Rumble-2-Flash speech recognition weights↩