SOV-Transcribe-1

SOV-Transcribe-1 is Salesforce’s speech-recognition model for turning recorded audio into text. Salesforce released it on 12 November 2026 with downloadable weights and integration examples for Transformers and vLLM.

It transcribes recordings in seven languages

The model card lists English, French, German, Spanish, Italian, Portuguese and Dutch. It accepts audio sampled at 16 kHz, with optional text instructions and context. The output is transcript text.

Salesforce describes a non-streaming system and reports testing recordings up to 30 minutes long. That makes its published configuration suitable for processing existing recordings. A real-time transcription product would need to account separately for how it receives, segments and processes incoming audio.

The weights use Apache 2.0

The complete model has 26.6 billion parameters, with 4.6 billion active, including its audio encoder. The launch article’s rounded 26B and 4B figures are rounded backbone counts. The model card provides the configuration to use when planning a deployment.

Apache 2.0 supplies the published licence terms. Running the model still requires suitable hardware, an inference runtime and a way to manage recorded audio. Those operational requirements are separate from the permission to download and use the weights.

Test the language and recording conditions you expect

Transcription accuracy depends on the language, accent, background sound and specialised vocabulary in a recording. Salesforce publishes its own evaluation results; teams should also test examples drawn from their intended use. Names, account references and important figures deserve particular attention when checking a transcript.

The opened sources establish downloadable weights and supported integrations. They do not establish a public hosted API price. Compare the cost of your own deployment with any hosted service’s current terms before estimating processing costs.

The cover is a conceptual illustration of transcription equipment.