Tasks and inputs
SOV-Transcribe-1 is listed for Automatic speech recognition, Speech → Text. It accepts audio, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Generic transcription models misspell the words your business depends on: product names, ticker symbols, the surname of the customer on the line. SOV-Transcribe-1, the first open-weight release from Salesforce Sovereign Intelligence (12 November 2026), lets you hand it those words in advance.
It is a speech-augmented language model for non-streaming speech recognition, released under Apache-2.0. Salesforce gives 26B total parameters with 4B active; the published repository lists 27B. Either way, it is a mixture-of-experts model that costs far less per token than its total size suggests.
Two controls set it apart from a classic recogniser.
Context biasing. Supply keywords, or a short description of the audio. On a contextual version of the Earnings-22 benchmark, keyword F1 rose from 71.9 to 96.8 when given the clip's own keywords, and reached 94.0 with the full keyword list. That second figure is the realistic one: you rarely know exactly which terms a clip contains.
Style instructions. By default it transcribes verbatim, ums included. A natural-language instruction can ask for digits instead of spelled-out numbers, or for disfluencies to be removed.
| Measure | Reported result |
|---|---|
| Word error rate (independent benchmark) | 2.28%; first among open-weight non-streaming models, seventh overall |
| VoiceCodeBench | 76.3% task success, second of 23 systems; 93.4% of entities recovered |
| Call-centre dialogues, 14 English accents | 8.21% average WER |
| Real earnings-call audio (HALAS) | 25.69% WER |
Those last two numbers are worth reading together. Clean benchmark audio and real earnings calls are very different worlds, and the gap is roughly tenfold. Publisher-reported throughout.
Salesforce says English is its strongest language and that others are supported, but it does not list them. It does not name the base model or the audio encoder, state a maximum audio length, or give serving instructions or hardware requirements. If you need any of those, test before committing.
Non-streaming means you send a finished recording. For live captions you need a different model, such as MAI-Transcribe-2-Streaming.
If your transcripts feed analytics where names and numbers matter (sales calls, earnings calls, support QA), the biasing feature is the reason to test. Pull 50 recordings where your current system got an entity wrong. Run them with no context, then with your standard keyword list, and count corrected entities as well as overall WER.
For comparison, look at Voxtral Mini Transcribe 2 and the full speech recognition list.
Compare the model on the work you need to complete.
SOV-Transcribe-1 is listed for Automatic speech recognition, Speech → Text. It accepts audio, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
ElevenLabs
Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.
NVIDIA
Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.
IBM
Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.
Mistral AI
Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.
Mistral AI
Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.
Mistral AI
Voxtral Small Transcribe 3 is Mistral's top-accuracy batch speech-to-text model for recorded audio, with context biasing and 50+ languages.