# Nvidia Nemotron 3.5 Transcribe: Open Speech Model Tested in 30+ Languages > Source: > Published: 2026-10-30 00:00:00+00:00 # Nvidia Nemotron 3.5 Transcribe: Open Speech Model Tested in 30+ Languages Nvidia's open-weight Nemotron 3.5 Transcribe handles 40+ languages, hour-long audio, and speaker labels. Here's how it performs, tested hands-on. ## What is Nvidia Nemotron 3.5 Transcribe? Nvidia Nemotron 3.5 Transcribe is an open-weight speech-to-text model that transcribes, translates, and diarizes audio across dozens of languages, including code-switched and low-quality recordings. It’s a 30 billion parameter mixture-of-experts model (released as the 30B-A4B variant) with roughly 4.3 billion parameters active per inference step, built on Nvidia’s Nemotron 3.5 architecture and fine-tuned from the Nemotron 3.5 Lightning 30B-A3B base model. It runs locally via tools like vLLM and outputs verbatim text, speaker tags, word-level timestamps, and translations, all from one model. ## TL;DR - **Nvidia released Nemotron 3.5 Transcribe** as an open-weight automatic speech recognition model on Hugging Face, built as a 30B-parameter mixture-of-experts model with about 4.3B active parameters per token. - **The model handles long audio natively** , processing recordings up to an hour long in one pass, with word-level timestamps and speaker diarization baked in. - **Language coverage is wide** , with tags on the model card spanning more than 35 languages including Arabic, Hindi, Japanese, Korean, Vietnamese, Thai, Turkish, and most of the major European languages, plus support for code-switching between languages mid-recording. - **A hands-on test across roughly 30 languages** from a real-world open dataset found strong results for widely-spoken languages and mixed results for some lower-resource or harder-to-transcribe ones, including noticeable errors in Urdu. - **The model is fast** , with Nvidia’s own benchmarks showing throughput above 2,000 seconds of audio processed per second under optimized quantization and speculative decoding, meaning a 30-second clip returns in a fraction of a second. - **Running it locally in full precision takes serious hardware** , around 75 GB of VRAM at runtime and about 67 GB of disk space, effectively requiring a single 80 GB GPU for production use, though lower-precision formats and smaller context windows reduce that footprint. - **The license is not fully open for commercial use** , since the model ships under an Nvidia OpenMDW-style license with specific terms that anyone deploying it commercially needs to check carefully. ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. ## How does Nemotron 3.5 Transcribe work? The architecture is built around a single instruction-following pipeline. You feed the model two things: audio and a plain-text instruction describing what you want back, whether that’s a verbatim transcript, a cleaned-up version, speaker-labeled dialogue, or a translation into another language. The model then produces that output directly. Under the hood, an audio encoder processes the raw sound and a speaker encoder distinguishes between different voices in the recording. Both feed into the 30 billion parameter language model backbone (Nemotron), which does the actual text generation. Because it’s a mixture-of-experts design, only a subset of the full parameter count (about 4.3 billion) is active for any given token, which keeps inference faster and cheaper than a dense model of equivalent size. A dedicated timestamp head aligns the generated text to timing information, so every word comes with a time code. This single-model, multi-output design means the same weights handle transcription, cleanup, diarization, and translation, with the task defined entirely by the instruction you pass in, rather than needing separate specialized models for each job. ## How well does it handle real-world, messy audio? One of the clearer demonstrations involved a phone call recording with two speakers talking over each other, full of filler words, interruptions, and unfinished sentences. The model transcribed every word, correctly labeled which speaker said what, and attached a timestamp to each word, all from a single pass through the audio. That kind of conversational, overlapping speech is traditionally one of the harder cases for ASR systems, so handling it cleanly in one shot is a meaningful signal of the model’s practical reliability rather than just benchmark performance on clean studio audio. ## How does it perform across different languages? A hands-on test ran the model across roughly 30 languages pulled from a real-world open dataset, deliberately including varied audio quality, dialects, and recording conditions rather than clean studio samples. Languages tested included Arabic, Bulgarian, Danish, German, Greek, Spanish, Estonian, Finnish, French, Hebrew, Hindi, Hungarian, Italian, Japanese, Korean, Lithuanian, Latvian, Maltese, Norwegian, Dutch, Polish, Portuguese, Romanian, Russian, Slovakian, Slovenian, Swedish, Thai, Turkish, Ukrainian, Urdu, and Vietnamese. Results varied by language and audio quality rather than falling into a simple pass/fail pattern. Major European languages (German, Spanish, Italian, French) came through clearly even with deliberately degraded audio. Harder cases like Hebrew and Japanese, tested with especially low audio quality and background noise, still produced largely usable output, though with more visible imperfections. Urdu stood out as a weaker result, with the transcription containing clear errors and disjointed phrasing that didn’t track well with the source audio. This spread suggests the model’s training data (Nvidia’s Granary-v2 dataset) likely has stronger coverage for some languages than others, which is typical for any multilingual ASR system but worth checking directly for your specific target languages before deploying it. ### Everyone else built a construction worker. We built the contractor. One file at a time. UI, API, database, deploy. The model card lists support for languages including Czech, Dutch, English, French, German, Hindi, Italian, Korean, Polish, Portuguese, Russian, Slovak, Spanish, Bulgarian, Croatian, Danish, Estonian, Finnish, Hungarian, Japanese, Latvian, Lithuanian, Romanian, Slovenian, Swedish, Thai, Turkish, Ukrainian, Arabic, Bengali, Chinese, Greek, Hebrew, Indonesian, Maltese, Marathi, Tagalog, Tamil, Telugu, and Vietnamese, putting total coverage above 35 languages with code-switching support for mixed-language speech within a single recording. ## How fast is it, and what hardware does it need? Nvidia’s own benchmarking shows the model can process more than 2,000 seconds of audio per second of compute time when run with optimized quantization formats like FP8 or NVFP4 combined with speculative decoding, compared to slower throughput at standard BF16 precision. In practical terms, a 30-second audio clip can come back in well under a second once the model is loaded and serving. That speed comes with a real hardware cost. Running the model in full precision consumes just over 75 GB of VRAM during inference and needs around 67 GB of disk space to store. That effectively points to a single GPU with 80 GB of VRAM, like an Nvidia A100 or H100, as the practical minimum for production deployment. VRAM usage can be brought down by reducing the context window or by switching to a lower-precision quantized format, but the full-precision version is squarely in the realm of data-center-grade hardware rather than a consumer GPU setup. ## Is Nemotron 3.5 Transcribe worth using? For teams that need multilingual transcription with diarization and timestamps out of one model, and that have access to high-VRAM GPU hardware, it’s a strong candidate. The combination of long-audio handling (up to an hour), instruction-driven flexibility (transcribe, clean up, translate, or label speakers from the same weights), and genuinely fast inference under quantization makes it practical for production pipelines rather than just a research demo. The caveats are real, though. The license is Nvidia’s own terms rather than a fully permissive open-source license, so anyone planning commercial use needs to read it closely before building a product around it. Language quality isn’t uniform, so lower-resource languages or unusual dialects deserve their own evaluation rather than assuming the strong results seen in major languages carry over everywhere. And the hardware floor (an 80 GB-class GPU for full precision) rules out casual local use on typical consumer graphics cards, even though quantized versions can narrow that gap. ## Frequently Asked Questions ### What is Nvidia Nemotron 3.5 Transcribe used for? It’s used for automatic speech recognition tasks: transcribing audio to text, translating spoken audio into other languages, labeling different speakers in a recording (diarization), and generating word-level timestamps, all controlled through natural-language instructions given alongside the audio. ### How many languages does Nemotron 3.5 Transcribe support? The model’s documentation lists support for more than 35 languages, spanning major European languages, several South and East Asian languages, Arabic, Hebrew, and others, with additional support for code-switching when speakers mix languages within the same recording. ### How much VRAM does Nemotron 3.5 Transcribe need to run? In full precision it uses just over 75 GB of VRAM at runtime and around 67 GB of disk space, pointing to a single 80 GB GPU for production deployment. Lower-precision quantized formats or a reduced context window can shrink that requirement. ### Is Nemotron 3.5 Transcribe free to use commercially? - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. It’s released as open-weight under an Nvidia license (not a standard permissive open-source license), so anyone considering commercial deployment should review the specific license terms on the model’s page before using it in a product. ### How does Nemotron 3.5 Transcribe compare on speed to other ASR models? Nvidia’s benchmarks show it can process more than 2,000 seconds of audio per second of compute under optimized settings (FP8 or NVFP4 quantization with speculative decoding), making short clips return in a fraction of a second once the model is loaded.