SOV-Transcribe-1

Salesforce released SOV-Transcribe-1 on November 12, 2026, establishing a new baseline for open-weight speech recognition. Built as a speech-augmented language model for non-streaming automatic speech recognition (ASR), the model blends an established audio frontend with a Mixture of Experts (MoE) text decoder. It achieves a 2.28% word error rate on the Artificial Analysis non-streaming benchmark, ranking first among open-weight speech models and outperforming comparable offerings from NVIDIA and Mistral.

The model is distributed on Hugging Face under a permissive Apache 2.0 license. Unlike pure transcription systems such as standard Whisper, SOV-Transcribe-1 uses its language model backbone to follow natural-language instructions, allowing engineers to bias recognition toward domain-specific terms, adjust formatting style, and maintain precision on critical entities like phone numbers, addresses, and code snippets.

For developers aiming to run SOV-Transcribe-1 locally, the primary hurdle is memory capacity. With 26.6B total parameters and 4.6B active parameters per token, it requires substantial VRAM to hold the entire weights table, even though its compute requirements during generation are remarkably low.

SOV-Transcribe-1 couples an audio perception frontend with an autoregressive MoE language model. The architecture consists of three distinct components:

• Audio Encoder: OpenAI's Whisper large-v3 encoder, handling 16 kHz audio inputs for recordings up to 30 minutes in length.
• Audio Projector: A trained cross-modal adapter that maps acoustic feature frames into the text embedding space.
• Decoder Backbone: A Google Gemma 4 26B A4B MoE language model operating in BF16, with the MoE router running in FP32 precision.

The MoE architecture dictates how the model behaves during inference. While 26.6 billion parameters must reside in memory, the network dynamically routes each token through only 4.6 billion active parameters. This separation creates a distinct profile for engineers sizing local systems: you pay the memory penalty of a 27B model to store all expert weights, but you enjoy the generation speed and throughput of a compact 4.6B model.

Salesforce also published a specialized DFlash draft model (SOV-Transcribe-1-DFlash) for speculative decoding. When paired with native vLLM serving, speculative decoding cuts inference latency significantly on longer transcriptions without degrading word error rates.

SOV-Transcribe-1 is built for high-stakes, non-streaming transcription where context and formatting dictate output quality. Supported languages include English, French, German, Spanish, Italian, Portuguese, and Dutch.

Standard ASR engines frequently corrupt alphanumeric strings, URLs, and proper nouns. SOV-Transcribe-1 scores near the top of benchmarks like VoiceCodeBench (#2 of 23 systems) specifically because its underlying language model understands structured data formats. It preserves email addresses, flight numbers, code identifiers, and telephone digits without phonetic drift.

Traditional ASR models require external post-processing pipelines to clean up spoken disfluencies or format numbers. With SOV-Transcribe-1, you specify your desired output style via system prompts:

• Verbatim Mode: Retains filler words ("um", "uh"), stuttered syllables, and exact pauses for legal depositions, compliance logs, or linguistic research.
• Cleaned-Up Mode: Removes filler tokens, fixes false starts, and outputs clean, publication-ready text for meeting notes or interview writeups.
• Formatting Control: Instructs the model directly on how to present numbers, currency, dates, and technical punctuation.

Engineers can supply prompt prefixes containing domain-specific glossaries, acronym sheets, or speaker metadata. The language model uses this prior context to disambiguate acoustically identical words, significantly lowering error rates on medical terminology, technical API names, and product catalogs.

Deploying this model locally requires planning for its split memory and compute profile. Because it is a local AI model with 26.6B parameters released in late 2026, memory bandwidth and capacity are the primary bottlenecks.

The base BF16 weights take up approximately 53 GB of storage. When loaded for execution with KV cache and runtime overhead, unquantized deployments require substantial GPU memory.

For single-GPU consumer deployments, Q4_K_M is the sweet spot. Because the Whisper audio encoder remains intact and the MoE router operates at full precision, 4-bit weight-only quantization on the Gemma MoE layers preserves transcription accuracy while dropping the total VRAM footprint below 20 GB. This makes it possible to run SOV-Transcribe-1 locally on a single 24GB workstation card.

• NVIDIA Consumer GPUs: If you want to know how to run a 26.6B model on consumer GPU setups, a single 24GB card (RTX 4090 or RTX 3090) works well using a 4-bit quantized build via vLLM or llama.cpp. To run BF16 without quantization degradation, dual RTX 3090 or 4090 cards running tensor parallelism are required.
• Apple Silicon: Macs equipped with unified memory handle the model cleanly. An M3 Max or M4 Max chip with 64GB or 128GB of memory can host the full BF16 or Q8 model weights alongside the context buffer with plenty of room to spare.
• Best GPU for SOV-Transcribe-1: For enterprise production deployments, an NVIDIA RTX 6000 Ada (48GB) running 8-bit quantization or a single 80GB A100/H100 PCIe card running unquantized BF16 provides the best GPU throughput and reliability.

Thanks to SOV-Transcribe-1 MoE efficiency, generation speeds are high once weights are loaded. Because only 4.6B parameters activate during decoding, you can expect 45 to 70 tokens per second on an RTX 4090 using 4-bit quantization. When running through vLLM with the official DFlash speculative decoding model enabled, token generation rates can exceed 90 tokens per second on high-bandwidth architectures.

Native execution is supported via:

1. vLLM: The recommended engine for batch throughput, utilizing the official SOV-Transcribe-1 vLLM plugin for non-streaming parallel inference.
2. Hugging Face Transformers: Useful for rapid prototyping and fine-tuning pipelines using AutoModelForMultimodalLM and AutoProcessor.

Evaluating SOV-Transcribe-1 against other modern speech models reveals clear architectural and deployment tradeoffs.

• SOV-Transcribe-1 vs NVIDIA Nemotron 3.5 Transcribe: Nemotron 3.5 Transcribe matches SOV-Transcribe-1 with roughly 4B active parameters. However, Salesforce's model achieves a 2.28% AA-WER compared to Nemotron's 2.81%. Nemotron uses a different parameter layout that can be marginally easier to serve on specific Triton configurations, but SOV-Transcribe-1 provides tighter instruction-following capabilities for entity-heavy audio.
• SOV-Transcribe-1 vs Mistral Voxtral Small: Voxtral Small relies on a 24B dense architecture (2.77% WER). Because Voxtral activates all 24B parameters on every token pass, its compute cost per token is much higher. SOV-Transcribe-1 requires similar total VRAM to host the 26.6B weights, but runs nearly three times faster during autoregressive decoding due to its 4.6B active MoE execution.
• SOV-Transcribe-1 vs Whisper Large v3: Whisper Large v3 is far lighter (1.5B parameters) and fits comfortably on almost any entry-level GPU with 8GB VRAM. However, it trails significantly in accuracy (4.62% WER on the same benchmarks), frequently hallucinates during long silences, and lacks native instruction steering.

If your hardware budget is limited to budget hardware with 8GB to 12GB VRAM, standard Whisper remains the default choice. If you have 24GB or more of memory and need unmatched transcription fidelity, entity extraction, and prompt-driven output styles, SOV-Transcribe-1 is the premier open-weight solution available.