~/SPEECH RECOG/salesforce-releases-sov-transcribe-1-open-weight-speech-augmented-language-model

Salesforce Releases SOV-Transcribe-1 Open-Weight Speech-Augmented Language Model

Salesforce has released SOV-Transcribe-1, an open-weight, speech-augmented language model designed for non-streaming automatic speech recognition with multilingual support and natural-language style control. The 26.6-parameter model ranks #1 among open-weight models for non-streaming speech-to-text on Artificial Analysis. This release provides developers and enterprises with a state-of-the-art, open-weight tool to handle complex transcription tasks while retaining full control over their audio data and infrastructure. It significantly advances open-source speech recognition, particularly in entity accuracy and natural-language style manipulation. SOV-Transcribe-1 features a Whisper large-v3 audio encoder combined with a Gemma 4 26B A4B MoE architecture, supporting up to 30 minutes of audio input. It is available under the Apache 2.0 license and supports inference frameworks such as vLLM and Hugging Face Transformers.

## BACKGROUND

Automatic speech recognition (ASR) is a technology that converts spoken language into written text, where non-streaming ASR processes the entire audio file at once instead of in real-time chunks. Speech-augmented language models integrate traditional audio encoders with large language models to better understand context and process speech inputs.

## KEYWORDS

#Speech Recognition#Open Source#AI Models#Salesforce#ASR

$ subscribe --daily