kaganerkan pixel logo

kaganerkan

News · Nov 16, 2026

SOV-Transcribe-1: Open-Weight Speech Recognition With Transcript Controls

Original five-ink diagram of an audio waveform becoming transcript lines beside formatting controls, labelled SOV-Transcribe-1 and Speech to Text.

Salesforce introduced SOV-Transcribe-1 on 12 November 2026, releasing the weights of a speech-recognition model that lets developers influence how recordings become text. The interesting change is not simply another accuracy claim: the model accepts instructions about the transcript’s format and context about difficult vocabulary. It is designed for non-streaming transcription, not advertised here as a live conversational system.

More control over the transcript

According to Salesforce’s launch article, the default output is a verbatim transcript. Instructions can instead request cleaner text, different number formatting or the removal of speech fillers. Context hints can supply names or domain-specific terms that would otherwise be difficult to recognize.

Those controls matter when a transcript feeds another application. An email address or account number can be wrong even when the surrounding sentence reads well. Formatting control is therefore a useful interface feature, not a guarantee that every important detail will be correct. Applications still need checks for the entities their workflow depends on.

Open weights do not mean a tiny deployment

The released model card lists the license as Apache 2.0. It lists 26.6 billion total parameters and 4.6 billion active parameters, including the audio encoder. The smaller active count should not be mistaken for the size of the complete model or a measured hardware requirement.

The model card describes audio inputs of up to 30 minutes tested, with optional text instructions and transcript text as output. It also says native support in vLLM and Hugging Face Transformers is being upstreamed; the released integration currently includes a vLLM plugin and repository-provided modeling code. That is a dependency to inspect before deployment, not a reason to execute remote code without review.

Running a model on infrastructure you control can change where audio is processed. It does not, by itself, establish a secure deployment, permission to process a recording or compliance with data-protection requirements.

The Turkish-language caveat

The model card lists English, French, German, Spanish, Italian, Portuguese and Dutch. Turkish is not listed. That does not prove that Turkish input will always fail, but this release documentation does not establish Turkish transcription support or quality.

Salesforce publishes benchmark results and ranking claims in its announcement. Those are publisher-reported results, not measurements reproduced for this article. In particular, an English result should not be presented as evidence of Turkish accuracy.

What to check before adopting it

For a transcription project, the next useful step is a small evaluation using recordings you are authorized to process: the required languages, noisy conditions, names, numbers and output formatting. Measure the full deployment’s resource needs and review its software dependencies. Treat corrections and human review as part of the workflow rather than assuming that open weights or a benchmark headline remove them.

Primary sources

Compiled with AI assistance from the linked primary sources. Publisher claims are attributed; no independent performance testing or personal verification by Kağan Erkan is claimed.