Salesforce opens speech-to-text model that can format finished transcripts

Salesforce released SOV-Transcribe-1 on November 12, giving developers a downloadable speech-to-text model for processing recordings on their own infrastructure. The model takes a completed audio file and returns a transcript. An optional text instruction can change how it writes numbers and punctuation or tell it to remove filler words. Salesforce published the weights on Hugging Face under the Apache 2.0 license.

The model card lists English, French, German, Spanish, Italian, Portuguese and Dutch. Korean is outside the listed language coverage. Salesforce tested recordings up to 30 minutes long. Teams can use the model for work that starts after a call or a meeting has been recorded. Live captions need a streaming system that can start producing text while someone is still talking. The server configuration documented by Salesforce caps each request at 10 minutes by default and can be set to accept up to 30 minutes.

One recording, different transcript formats

The default instruction is to transcribe speech verbatim. Users can also specify how the output should handle capitalization, punctuation and numbers. A team that needs to preserve the order and wording of a call can request a verbatim transcript. A dictation tool can send the same recording with instructions to remove hesitations and render spoken numbers as digits. The instructions shape the transcript as the model writes it.

The model can take a short description of the recording or a list of expected terms. This gives it context for names, company names and specialist vocabulary that may sound like more common words. Salesforce tested that feature on clips from corporate earnings calls. The model's keyword F1 score rose from 71.9 to 96.8 when the team supplied the relevant terms for each clip. That score measures how reliably the model recovers the target terms and avoids false matches in that test.

What the 2.3% error rate measures

Artificial Analysis lists a 2.3% word error rate for SOV-Transcribe-1 on its non-streaming benchmark. It was the lowest rate among downloadable-weight models in the evaluator’s November 12 comparison. The evaluator counts substitutions, missing words and added words against a verified transcript, then divides the total by the number of words in that transcript. A single incorrect character can still make an email address unusable, even if the rest of the transcript reads correctly.

The benchmark combines roughly eight hours of English audio. Conversations with voice agents account for half the index. European Parliament speech and company earnings calls each account for a quarter. The score measures accuracy on that mixture of recordings, with their particular speakers and audio conditions. Artificial Analysis evaluates streaming speech separately and measures how long it takes to get both partial and final text.

Salesforce reported different error rates in two other tests. It measured 8.21% across customer-service conversations covering 14 English accents and 25.69% on a set of earnings-call recordings selected to test whether systems add words that were never spoken. These were separate tests published by the developer. The differences show how much the result can depend on the recording and the test set. The independent English score covers the particular samples in its benchmark.

Salesforce combined a Whisper audio encoder with a Gemma 4 language model. Its model card specifies 26.6 billion parameters in total, of which 4.6 billion are active during processing, including the audio encoder in the count. The active count describes how much of the model contributes to a given computation. An organization deploying the model still needs infrastructure that can store and load the full weights. Hugging Face lists the repository at about 53.3 GB. That figure describes file storage. Running the model also takes memory for the audio input and the data it generates during processing.

The release includes model code and weights. Salesforce documented deployment with Hugging Face Transformers and vLLM using code or a plugin supplied with the release. Native support in those tools was still being incorporated when the model card was published. The launch post lists a separate technical report as forthcoming.