NVIDIA Nemotron 3.5 Transcribe: 5.33% WER Multilingual ASR
By Ali Al-Montaser — Senior AI Technical Editor
NVIDIA released Nemotron 3.5 Transcribe, an open-weight ASR model on Hugging Face and ModelScope, topping the English Open ASR leaderboard at 5.33% WER. It unifies transcription, speaker diarization, translation, and audio QA in one checkpoint, supports 40 languages, and ships with the 463,062-hour Granary v2 dataset.
Executive Overview
NVIDIA has officially released Nemotron 3.5 Transcribe, an open-weight speech model available on Hugging Face and ModelScope. It unifies transcription, speaker diarization, translation, and audio question answering in a single checkpoint. The model tops the Hugging Face Open ASR Leaderboard for English with a 5.33% word error rate (WER) and ships alongside the open Granary v2 dataset, which spans 463,062 hours across 48 languages. This release positions NVIDIA at the forefront of open, multilingual speech AI, offering enterprises a free, high-accuracy alternative to proprietary APIs.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Open-weight model, free to download and use from Hugging Face and ModelScope; no token pricing announced in the official statement. Actual cost depends on GPU hosting (e.g., RTX PRO 6000 Blackwell Server Edition) or inference via vLLM. |
| 🌐 Platforms & Immediate Availability | Hugging Face and ModelScope (available now); serving via vLLM (available today); local inference coming via NeMo-Speech.cpp and quantized versions on RTX Spark, DGX Spark, Jetson AGX Thor, and Blackwell data center modules. |
| ⚡ Performance & Speed Benchmarks | 5.33% English WER (first among open models); 4.62% multilingual WER; 10.98% long-form WER; 2.8% AA-WER; 1.8% WER on AA-AgentTalk; 18.26% DER on 21 languages; 32.4% cpWER on VoiceArena; processes 100 hours of audio in under 3 minutes on a single RTX PRO 6000 Blackwell (NVFP4 version). |
| 🛡️ Security & Robustness | No explicit security standards or prompt injection resistance mentioned in the statement; Granary v2 includes hallucination check in ASR stage and a per-record quality flag. |
| 🧠 Context Window | Processes audio recordings up to one full hour (60 minutes) in a single pass, with word-level timestamps support. |
| 🌍 Arabic & Regional Support | 40 transcription languages, 31 translation languages to/from English, 13 speaker diarization languages; Arabic is not explicitly specified in the statement and must be verified via the model card and per-language task matrix. |
Deep-Dive Features & Architecture
Nemotron 3.5 Transcribe is built on a hybrid architecture that couples a speech encoder with a Mixture-of-Experts language model combining Mamba and Transformer layers. It activates only about 4 billion parameters out of 30 billion per token, placing it on the Pareto frontier tracked by Artificial Analysis for accuracy versus active parameters. This design outperforms widely used commercial speech APIs on the comprehensive AA chart. The model executes multiple tasks from a single checkpoint via natural language prompts: transcription with speaker labels, translation to English, or answering questions about a meeting, with support for medical vocabulary steering (e.g., HbA1c, metformin) within the text prompt.
On the data side, Granary v2 provides a reusable curation pipeline built on NVIDIA NeMo Curator, covering 463,062 hours across 48 languages and 16 source datasets (up from 25 languages in the first version). The pipeline produces three text forms per recording (tn_raw, itn_raw, and itn_clean), plus task labels, short audio captions, code-switching recovery, audio-linked QA pairs, contextual ASR terms, and translations with automatic quality scores. In the High Fidelity stage, VoxPopuli data was reduced from 354,000 hours to 111,000 hours across 22 European languages.
Benchmark & Competitive Performance
On the Hugging Face Open ASR Leaderboard, the model ranks first among open-weight models in English with 5.33% WER, and scores 4.62% in multilingual and 10.98% in long-form tracks. On Artificial Analysis, it ties for first place among open models at 2.8% AA-WER with about 4 billion active parameters. Additional metrics include 1.8% WER on AA-AgentTalk, 18.26% DER on 21 languages, and 32.4% cpWER on VoiceArena. The model processes 100 hours of audio in under 3 minutes on a single RTX PRO 6000 Blackwell GPU (NVFP4 version), demonstrating exceptional throughput for enterprise-scale transcription.
Industry Impact & Enterprise Adoption
The release of Nemotron 3.5 Transcribe and Granary v2 lowers the barrier to entry for organizations seeking high-accuracy, multilingual speech AI without proprietary API lock-in. The open-weight license allows fine-tuning and on-premises deployment, addressing data sovereignty and latency concerns. NVIDIA's provision of fine-tuning recipes that reduced WER by 45.8% in a medical study underscores the model's adaptability to domain-specific vocabularies. With support for 40 transcription languages, 31 translation languages, and 13 diarization languages, the model is poised to accelerate adoption in healthcare, media, customer service, and legal transcription. The availability of quantized versions for edge devices (RTX Spark, Jetson AGX Thor) further extends its reach to real-time and offline applications.
Conclusion
NVIDIA's Nemotron 3.5 Transcribe sets a new benchmark for open-weight speech recognition, combining state-of-the-art accuracy with a unified multi-task architecture and a massive open dataset. Its top ranking on Hugging Face Open ASR and competitive AA-WER scores make it a compelling choice for developers and enterprises alike. As the model becomes available across platforms and edge devices, it is expected to drive innovation in multilingual voice interfaces and audio analytics. For teams evaluating speech AI, Nemotron 3.5 Transcribe warrants immediate attention.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Nemotron 3.5 Transcribe achieves 5.33% average WER on the English track of the Hugging Face Open ASR Leaderboard (first among open-weight models), 4.62% on the multilingual track, and 10.98% on the long-form track. On Artificial Analysis, it scores 2.8% AA-WER and 1.8% on AA-AgentTalk.
It supports transcription in 40 languages, translation to and from English in 31 languages, and speaker diarization in 13 languages. It can process recordings up to one hour long and identify up to 8 speakers.
It is available now on Hugging Face and ModelScope as open weights. It can be served via vLLM today, with upcoming support for local inference via NeMo-Speech.cpp and quantized versions on RTX Spark, DGX Spark, Jetson AGX Thor, and Blackwell data center modules.
Granary v2 is an open data pipeline for multilingual audio curation and preparation, covering 463,062 hours across 48 languages and 16 source datasets (up from 25 languages in the first version). It produces three text forms per recording: tn_raw, itn_raw, and itn_clean.
The announcement does not explicitly list Arabic among the 40 transcription languages, 31 translation languages, or 13 diarization languages. Users should check the model card and per-language task matrix on Hugging Face to verify Arabic support before relying on it in production.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
