Salesforce Releases SOV-Transcribe-1 Speech Model

Salesforce has released SOV-Transcribe-1, an open-weight speech-to-text model that the company says ranks first among open-weight non-streaming models on Artificial Analysis, and seventh overall, with model inclusion and rankings as of 12 November 2026. The weights are being published on Hugging Face under an Apache 2.0 licence. It is a non-streaming automatic speech recognition model, not a claim to be the best speech model overall.

The launch blog, published on 12 November 2026, describes SOV-Transcribe-1 as Salesforce’s “first open-weight speech-augmented language model, with 4 billion active parameters (26 billion total), built for non-streaming automatic speech recognition (ASR).” The model card gives a more precise figure: 26.6 billion total parameters and 4.6 billion active, including the audio encoder. The blog rounds those numbers.

According to the card, the architecture pairs a Whisper large-v3 audio encoder with a Gemma 4 26B A4B mixture-of-experts language model, plus an audio projector. Salesforce says the model has been tested on audio of up to 30 minutes, and that inference is supported in vLLM and Hugging Face Transformers. The card lists seven languages: English, French, German, Spanish, Italian, Portuguese and Dutch.

Rankings and figures Salesforce reports

Every ranking and figure below is Salesforce’s, as of 12 November 2026. TSN has not independently scored the model.

On Artificial Analysis, Salesforce says SOV-Transcribe-1 ranks “#1 among open-weight models for non-streaming speech-to-text on Artificial Analysis (#7 overall).” It reports an Artificial Analysis word error rate (AA-WER) of 2.28 per cent, against 2.81 per cent for NVIDIA’s Nemotron 3.5 Transcribe and 2.77 per cent for Mistral’s Voxtral Small. The blog presents that result as beating Nemotron 3.5 Transcribe at the same number of active parameters, using the rounded 4 billion figure, and Voxtral Small with fewer active parameters.

On VoiceCodeBench, which BeSimple runs as a private leaderboard of whether a system captures key terms such as phone numbers, email addresses, URLs and code symbols, Salesforce says SOV-Transcribe-1 ranks second of 23 systems overall, and first among open-weight models, on both of the measures it cites. It reports that 76.3 per cent of recordings pass — a recording passes only if every target entity in it is correct — and that 93.4 per cent of entities are recovered.

On AppTek Call-Center Dialogues, a long-form benchmark of agent–customer calls across 14 English accents, Salesforce reports an 8.21 per cent average word error rate, and the lowest word error rate on eight of the 14 accents: Australian, British, Scottish, Welsh, Irish, Indian, Mexican and Singaporean English.

On HALAS, a human-annotated benchmark of real earnings-call audio that Salesforce says models are most likely to hallucinate on, the company reports a 25.69 per cent word error rate, which it says is lower than other open-weight models it lists, including Alibaba’s Qwen3-ASR 1.7B at 26.01 per cent, Mistral’s Voxtral Small 24B at 27.35 per cent, NVIDIA’s Nemotron 3.5 Transcribe at 31.59 per cent, and IBM’s Granite Speech 5.0 2B at 34.51 per cent.

Salesforce also reports a context-biasing test, Contextual Earnings-22, on 15-second clips of earnings calls built around the name of a person, company or product. Given the clip’s keywords, it says the model raises its keyword F1 score — a measure of how accurately those names are transcribed — from 71.9 to 96.8. Given the full hour-long call’s keyword list, most of them distractors, it still scores 94.0, which Salesforce says is the highest of the open-weight models it compared.

The blog describes further controls that are not rankings. The model is instruction-controlled: it transcribes verbatim by default and can change formatting from a natural-language instruction, such as writing numbers as digits or removing disfluencies. It also accepts keywords or a short description of the audio to bias recognition toward domain terms. A technical report is described as coming soon.

TSN recently covered another speech-to-text launch, ElevenLabs’ Scribe v3 and Realtime Scribe v3 Live. Salesforce does not compare SOV-Transcribe-1 with Scribe v3 in the figures above, and that release is a separate, commercial speech product rather than a like-for-like open-weight ranking.

Analysis

The headline ranking is narrower than a casual reading of “#1” might suggest. Salesforce’s own wording places SOV-Transcribe-1 first among open-weight non-streaming speech-to-text models on Artificial Analysis, and seventh on that leaderboard overall, as of 12 November 2026. It is not, on that evidence, the best speech model available, and TSN should not be read as saying that it is. Model inclusion can change, and these are company-reported standings, not a TSN score.

The parameter count needs the same care. The launch blog’s “4 billion active” and “26 billion total” are rounded. The model card’s 26.6 billion total and 4.6 billion active, including the audio encoder, are the figures to use when describing the model’s size. Salesforce’s comparison with Nemotron 3.5 Transcribe “with the same number of active parameters” rests on that rounded 4 billion figure, not on a TSN check of both cards.

Entity accuracy and average word error rate are also different claims. A 2.28 per cent AA-WER can still hide a swapped digit in a phone number or a dropped character in an email address, which is why Salesforce leans on VoiceCodeBench’s stricter pass rate: 76.3 per cent of recordings pass only when every target entity is correct, even though 93.4 per cent of entities are recovered. Those are useful caveats for anyone planning to put transcripts into tools or databases. The HALAS and AppTek figures, likewise, are Salesforce’s results on noisy earnings calls and accented call-centre audio, not a general guarantee that the model will hold up in every real-world channel.

The release is also limited in ways the marketing language can skip. It is a non-streaming model, so it is not being offered here as a live captioning stack. The card lists seven languages, not a claim of dozens. And “tested” on recordings of up to 30 minutes is not the same as a hard product limit published with a full technical report; Salesforce says that report is still to come. For developers who want to run speech recognition on infrastructure they control, an Apache 2.0 weight release is the practical point, provided they judge the company’s numbers against their own audio.