Introducing Salesforce SOV-Transcribe-1: #1 Open-Weight Model Built for Real-World Speech
Blog post from Hugging Face
Salesforce has introduced SOV-Transcribe-1, an Apache 2.0-licensed open-weight, non-streaming speech-to-text model with 4 billion active parameters and 26 billion total parameters, intended for enterprise-controlled voice applications such as agents, dictation, and call transcription. The company reports that the model leads open-weight systems on several benchmarks, ranking first for non-streaming transcription on Artificial Analysis and for entity-sensitive performance on VoiceCodeBench, where accurate handling of phone numbers, email addresses, URLs, and symbols is essential. Salesforce also highlights its reported performance across 14 English accents, noisy real-world earnings-call audio, and instruction-based formatting controls that can preserve verbatim speech or remove fillers and format numbers. The model supports context biasing through supplied keywords or audio descriptions, which Salesforce says substantially improves transcription of difficult names, companies, products, and domain-specific terms. Model weights are available on Hugging Face, while a technical report is planned.
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.