On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an AI model that transcribes speech in 40 languages. Its defining feature is that users can specify how numbers should be written and which technical terms to use, so the same audio can yield different transcripts depending on the purpose. In addition to standard speech recognition, it supports speaker-attributed transcription and translation between English and other languages. According to the official model card, Japanese is among the supported languages.
In its technical report, NVIDIA says the model recorded an average word error rate (WER) of 4.71% across nine public English datasets, ahead of the open-weight models it was compared against. That figure alone, however, does not tell us how accurate the model is in Japanese or how well it handles specialized terminology. When comparing it with existing models such as Qwen and Whisper, transcription accuracy is only part of the picture: adherence to specified notation, speaker separation, and the computing resources needed to run it in practice all matter as well.
A Speech Recognition Model That Accepts Instructions on Terminology and Formatting
Nemotron 3.5 Transcribe is a speech recognition model that can process recordings of up to one hour in a single pass. It supports transcription in 40 languages, two-way translation with English in 31 languages, and speaker-attributed transcription in 13 languages, handling up to eight speakers. Not every feature is available in all 40 languages. Input audio is 16 kHz mono, and the output is text. The supported scope for each task can be checked in NVIDIA's announcement and the official model card.
What sets it apart is that users can instruct not only the language to be recognized but also the finish of the transcript. For example, you can choose between a faithful verbatim record that keeps disfluencies, or a cleaner, more readable version with normalized numbers and disfluencies removed. Users can also supply candidate technical terms and proper nouns along with their preferred spellings, so the same recording can produce output suited to different uses. Still, being able to control the transcript format is not the same as being able to understand the content of a meeting and summarize it accurately.
This approach differs from simply running a separate language model over speech recognition output to reformat it. According to NVIDIA's technical report (Figure 2 and Chapter 2), a speech recognizer of about 600 million parameters runs in parallel with "Sortformer," a component of about 100 million parameters whose weights are frozen. Sortformer handles speaker diarization, estimating who is speaking and when. The acoustic features extracted by the recognizer are combined with time-based speaker activity information and passed to the language model, which generates text that reflects both what was said and who said it. In other words, speaker information is used from the generation stage rather than being attached as labels after the text is produced.
Internally, to make efficient use of a diarizer trained on short audio segments, the system extracts audio features in 30-second units. The language model then joins the features from each segment and keeps speaker labels consistent across long recordings. This is not a mechanism that returns recognition results in real time every 30 seconds. After text generation, a timestamp adapter based on CTC (Connectionist Temporal Classification) aligns the audio with the text and produces word-level timestamps. Note that speaker labels are for distinguishing speakers, not for identifying individuals.
The ability to change notation through instructions is also linked to how the training data is structured. "Granary v2," released by NVIDIA, provides multiple reference transcripts for the same utterance: a verbatim version with disfluencies, a version with normalized notation such as numbers while retaining disfluencies, and a version with disfluencies removed. NVIDIA says it performed supervised learning on these instruction-and-reference pairs and then added a short period of reinforcement learning to improve recognition accuracy. The model is thus trained not just to convert speech into characters but also to produce notation that fits the stated purpose.
Accuracy Compared With Qwen and Whisper: 4.71% Average English WER
Nemotron 3.5 Transcribe recorded an average WER of 4.71% across nine public English datasets. WER is the number of word substitutions, deletions, and insertions relative to the reference text, divided by the number of words in the reference. A lower value means fewer transcription errors. Comparing Tables 6, 7, and 9 of NVIDIA's technical report shows how the model stands against open-weight models and where it sits once commercial services are included.
| Model | English, 9-set average WER | Long-form, 3-set average WER | Multilingual, 7-language average WER |
|---|---|---|---|
| Nemotron 3.5 Transcribe | 4.71 | 10.98 | 4.62 |
| Qwen3-ASR-1.7B | 4.94 | Not listed | 6.49 |
| Granite-Speech-5.0-2B | 5.06 | Not listed | Not listed |
| Canary-Qwen-2.5B | 5.38 | 14.08 | Not listed |
| Parakeet-TDT-0.6B-v3 | 5.58 | 13.37 | Not listed |
| Voxtral-Small-24B (2507 version) | 5.81 | Not listed | 6.52 |
| Whisper-large-v3 | 6.66 | 13.92 | 7.94 |
| OpenAI gpt-transcribe (as labeled in the technical report) | 5.49 | Not listed | Not listed |
| ElevenLabs Scribe v2 | 4.55 | 9.05 | 3.56 |
| OtterAI speech-v3.1 | 3.88 | Not listed | Not listed |
| AssemblyAI Universal-3 Pro | Not listed | 10.35 | Not listed |
All figures are percentages. "Not listed" means no corresponding comparison result is published; it does not indicate that the feature is unsupported.
The English results are a simple average over nine datasets, with comparison values retrieved on October 21, 2026. The long-form results cover three datasets, Earnings21, Earnings22, and CORAAL, using regional averages for CORAAL. They were evaluated by splitting recordings into 55-minute segments with a 5-second overlap between adjacent segments. The multilingual results are a simple average of per-language averages for seven languages: German, French, Italian, Spanish, Portuguese, Dutch, and Hindi. Japanese is not included, and the multilingual comparison values were retrieved on October 2, 2026.
Among open-weight models, Nemotron recorded the lowest average English WER. On long-form audio it also performed well, at 10.98% against 13.92% for Whisper-large-v3. Widening the comparison to commercial services, however, ElevenLabs Scribe v2 recorded lower WERs than Nemotron across the board: 4.55% for English, 9.05% for long-form, and 3.56% for multilingual. Doing well among open-weight models is not the same as outperforming every speech recognition service.
On the nine-dataset English average, the gap between Nemotron's 4.71% and Qwen3-ASR's 4.94% is 0.23 points. Taking Qwen3-ASR as the baseline, the relative error reduction works out to about 4.7% (0.23 ÷ 4.94 × 100). This is calculated from the comparison values NVIDIA reported, however, and is not an exhaustive comparison with prompts and inference conditions fully unified across all models. That gap alone cannot establish statistical significance or tell us which model is better in terms of operating cost. The technical report was also written by NVIDIA researchers, and there is no confirmation that it has been peer reviewed.
On individual datasets, competing models sometimes beat Nemotron. On the meeting-speech set AMI, Granite-Speech scored 6.96% against Nemotron's 7.81%. On URGENT24, which includes degraded audio, Qwen3-ASR scored 9.93% against Nemotron's 11.00%. Average WER alone cannot show which model is better in a particular recording environment. For recordings with technical terms or multiple speakers, how far the output can be controlled is an important comparison point alongside raw recognition accuracy.
Specifying Terminology Sharply Cuts Error Rates; Strength in Speaker Diarization
NVIDIA has published concrete evaluation results for the feature that lets users specify technical terms and proper nouns in advance. According to Table 11 of the technical report, on ContextASR-Bench's "Speech English," overall WER was 3.33% and NE-WER, which measures the error rate on named entities, was 12.21% when no terms were specified. With detailed term specification added, overall WER fell to 1.71% and NE-WER to 2.20%. The named-entity error rate improved dramatically, confirming the benefit of supplying appropriate term information.
However, the results vary greatly depending on how the terms are specified. With vague term specification, overall WER was 3.34% and NE-WER was 12.26%, barely any improvement over specifying nothing. In other words, even though the model can accept terminology instructions, it cannot necessarily fill in the correct spellings automatically from loose guidance. In real use, the relevant proper nouns and technical terms need to be specified concretely.
Nor does Nemotron lead on every metric against competing models. Under the same detailed term specification, Qwen3-ASR scored 1.53% overall WER and 2.76% NE-WER. Nemotron was better on named-entity accuracy, but Qwen3-ASR had the lower overall WER. Qwen3-Omni-30B-A3B recorded 1.13% overall WER and 1.55% NE-WER. These Qwen-family figures are from additional internal measurements by NVIDIA, and Nemotron's large improvement cannot simply be read as an absolute advantage in recognition accuracy.
"Preference-ASR," which evaluates adherence to specified notation, showed a different pattern. Nemotron's WER fell from 7.56% without instructions to 6.16% with them. Canary-Qwen, by contrast, worsened from 8.16% to 8.36%, and Qwen3-Omni from 8.30% to 10.34%. Unlike ordinary transcription, this evaluation also scores whether the output follows the specified notation. In environments where a glossary can be prepared and notation rules for minutes or technical documents are settled, Nemotron's instruction-following ability is worth evaluating.
For multi-speaker speech recognition, the evaluation uses "cpWER," which assesses the accuracy of the content and its assignment to speakers together. This metric optimizes the correspondence between estimated speaker labels and the reference speakers, then computes the word error rate. Unlike ordinary WER, it evaluates not only the transcript content but also whose speech each part was recorded as. The average results over six conditions with the same audio and scoring conditions, as listed in Table 14 of the technical report, are as follows.
| Method | Average cpWER (%) |
|---|---|
| Nemotron 3.5 Transcribe | 15.80 |
| DiCoW v3.3 + DiariZen-v2 | 16.83 |
| Multi-talker Parakeet + Nemotron 3 Diarization | 18.71 |
| MOSS-Transcribe-Diarize | 19.26 |
| VibeVoice-ASR | 24.25 |
Averaged over the six conditions, Nemotron recorded the lowest cpWER. In individual conditions, though, competitors came out ahead. On NOTSOFAR1, recorded with a single-channel microphone placed at a distance, Nemotron scored 25.91% against DiCoW's 24.80%. On Mixer6, MOSS recorded 13.25% against Nemotron's 14.80%.
Microphone conditions also have a large effect. Nemotron's cpWER was 10.29% under the IHM condition, which mixes audio from each meeting participant's headset microphone, but worsened to 17.34% under the SDM condition, which uses a single distant microphone. The approach of processing speech content and speaker information together is effective, but it does not eliminate differences in accuracy caused by the recording environment. For use in settings such as meeting rooms, evaluation that reproduces the actual microphone placement and the way speakers overlap is essential.
Japanese Accuracy and Third-Party Benchmark Results
In addition to Japanese transcription, Nemotron 3.5 Transcribe supports translation to and from English, speaker-attributed recognition, and speech in which multiple languages are mixed. NVIDIA, however, classifies Japanese as "Production-ready" rather than in the top "Leading" tier. According to the Japanese evaluation in the official model card, the character error rate (CER) was 4.42% on FLEURS and 10.61% on the CoVoST2 speech recognition task. The latter is a figure for transcription accuracy, not an evaluation of translation quality.
CER is an error rate per character and has a different denominator from WER, which is calculated per word. The Japanese CER of 4.42% therefore cannot be compared directly with the English average WER of 4.71% or the seven-language average WER of 4.62%, which excludes Japanese. We also could not find any independent evaluation comparing competing models on Japanese under the same conditions. Japanese is certainly supported, but there is no basis for concluding that this is the most accurate model for Japanese transcription.
Among evaluations by parties other than the developer is Artificial Analysis's non-streaming speech recognition benchmark. Its AA-WER uses about eight hours of audio and is computed as a time-weighted average, with AgentTalk at 50% and VoxPopuli-Cleaned-AA and Earnings22-Cleaned-AA at 25% each. In the open-weight model ranking checked on October 28, 2026, Voxtral Small was first and Nemotron second. Both models' displayed WER was 2.8%, but because the figures are rounded, this does not mean they are strictly tied for first place.
Another independent evaluation is Bluejay Labs' MISTT. This benchmark uses 2,999 clips, including voice agent calls, doctor-patient conversations, and parliamentary speech in five languages, plus 400 degraded audio files generated from 100 recordings. In the English evaluation, Pulse Pro recorded a WER of 3.38% (95% confidence interval 3.10–3.66%) and Amazon Transcribe 3.43% (3.12–3.72%). Nemotron scored 3.54% (3.23–3.86%) and ElevenLabs Scribe v2 3.60% (3.26–3.97%).
The confidence intervals were calculated using the bootstrap method with 1,000 resamples, and the models' intervals overlap. It is therefore difficult to conclude, from point estimates of WER alone, that the ranking among the models is statistically established. The types of audio and weighting also differ among Artificial Analysis, MISTT, and NVIDIA's nine public datasets. Nemotron's figures of 2.8%, 3.54%, and 4.71% should not simply be lined up and read as the same model's accuracy varying under identical conditions.
It is also necessary to distinguish transcription accuracy from the ability to understand audio content and answer questions about it. NVIDIA's evaluations of audio understanding, reasoning, and following spoken instructions cover English only. On the non-music questions of MMAU-Pro, Nemotron's accuracy was 52.7%, compared with 57.5% for Qwen3-Omni. Question-answering about audio is also offered by models such as Voxtral. The fact that Japanese transcription is supported does not mean that comparable performance can be expected in Japanese audio understanding or reasoning.
Up to 2,238x Real-Time Throughput: Measurement Conditions and Required GPUs
NVIDIA has published a throughput figure for Nemotron 3.5 Transcribe of up to 2,238 RTFx. RTFx indicates how many times longer than the actual processing time the audio is that can be processed. At 2,238 RTFx, the system can process a total of 2,238 seconds of audio for every second of real time. However, this is aggregate throughput when processing many audio files simultaneously, not the response time for a single recording.
According to Section 4.13 of the technical report, this peak figure was obtained using an "RTX PRO 6000 Blackwell Server Edition" with 96 GB of memory, combining NVIDIA NIM optimization with NVFP4 quantization. It also uses "DFlash2," a speculative decoding technique that looks ahead at candidates for upcoming tokens and verifies them with the main model. The measurement processed 30-second recordings at high concurrency and did not request word-level timestamps. This value therefore cannot be treated as the throughput obtainable on typical GPUs or default settings.
NVIDIA also reports a result of processing a single 30-second recording in 0.13 seconds. That figure was obtained under a different load condition from the 2,238 RTFx peak, and the two cannot be treated as measurements under the same conditions. Moreover, when the recording extends to 10 minutes, the technique that yields maximum throughput changes from DFlash2 to "MTP (Multi-Token Prediction)," which predicts multiple tokens. Actual throughput depends on recording length, the number of concurrent requests, and the decoding technique used.
Importantly, these optimizations were not integrated into the NeMo and vLLM implementations at the time of release. NVIDIA expects the released developer-facing implementations to run at roughly half the speed of the optimized environment. The 2,238 RTFx figure is not a result reproduced by third parties using the public implementation. To judge the performance you can expect on deployment, you need to measure it in the software environment actually available.
Under other conditions listed in the model card, batch processing on an H100 with BF16 is given as roughly 410 RTFx for MTP2, about 270 RTFx for MTP8, about 340 RTFx for DFlash2 with K7, and about 260 RTFx for K15. The number after MTP and the number after K in DFlash2 indicate how many candidate tokens are looked ahead. All of these figures exclude startup time and represent throughput, not single-request response time. These results also show that increasing the number of lookahead tokens does not always make processing faster.
Regarding GPU memory, it is best not to judge the required capacity from the "30B-A4B" in the model name alone. Nemotron 3.5 Transcribe has about 34 billion total parameters, of which about 4.3 billion are active during text generation. Even with a small number of active parameters, the full set of model weights must be held during inference, so it will not necessarily run in memory comparable to a small model of around 4 billion parameters. In particular, the required memory varies with the quantization format.
The official model card lists an H100 for the standard BF16 configuration, a 48 GB L40S or H100 for FP8, and an RTX 5090 for NVFP4. Reducing memory use through quantization, however, involves a trade-off with recognition accuracy. According to Section 4.12 of the technical report, across 77 evaluation sets on a single GB300, the average relative increase in error rate from quantization was 1.0% for FP8 and 3.6% for NVFP4. Some evaluation items, including non-English and speaker-attributed recognition, show lower accuracy, so similar average English performance does not mean the same results for every use. The quantization format should be chosen on the basis of both memory capacity and recognition accuracy, taking the languages and recording conditions into account.
How Nemotron Differs From Qwen, Parakeet, and Voxtral, and How to Choose by Use Case
When choosing a speech recognition model, processing method, supported languages, and available computing resources matter as well as WER rankings. For example, Qwen3-ASR supports both batch processing of recordings and streaming recognition. Nemotron 3.5 Transcribe, by contrast, is distinguished by its use with pre-recorded audio, where notation, terminology, and speaker-attributed output can be specified.
Organizing each model's specifications on the basis of the official model cards for Nemotron, Qwen3-ASR, Parakeet, and Voxtral gives the following.
| Model | Supported languages and main features | Model size and license |
|---|---|---|
| Nemotron 3.5 Transcribe | 40 languages, batch processing of recordings, speaker-attributed recognition in 13 languages, notation control via instructions, Japanese supported | 34B total, about 4.3B active, OpenMDW-1.1 |
| Qwen3-ASR-1.7B | 30 languages plus 22 Chinese dialects, batch processing and streaming, Japanese supported, timestamps via a separate model | 1.7B, plus a separate 0.6B model for timestamps, Apache 2.0 |
| Parakeet-TDT-0.6B-v3 | 25 European languages, timestamped recognition, suited to large-volume processing, Japanese not supported | 0.6B, CC BY 4.0 |
| Voxtral-Small-24B-2507 | 8 languages, Japanese not supported, up to 30 minutes for transcription and up to 40 minutes for audio understanding, question answering and summarization | 24B, 32k context, Apache 2.0 |
To obtain word-level timestamps with Qwen3-ASR, you separately use "Qwen3-ForcedAligner-0.6B," which supports 11 languages including Japanese.
When recognition results must be returned in real time, Qwen3-ASR, which supports streaming, is a candidate for comparison. For Japanese recordings, Parakeet-TDT-0.6B-v3 and Voxtral-Small-24B-2507, whose supported languages do not include Japanese, are not candidates for the same use. Where the supported languages do match, there are options such as using the compact 0.6B Parakeet for timestamped transcription, or using Voxtral to handle question answering about audio as well. Parameter counts do not directly indicate actual GPU memory use or operating costs, and license terms differ by model, so these should be checked against the intended deployment.
NVIDIA has also released datasets that can be used for additional training in specific domains. For Granary v2, the published label listing data spans 463,062 hours, 48 languages, and 16 source corpora. This is not a release of all of Nemotron's training data, however, and the original audio must be obtained separately under each source's license terms. Data that provides multiple notations for the same utterance is useful not only for improving recognition accuracy but also for teaching output formats suited to different purposes.
As an example of the effect of additional training, NVIDIA reports an adaptation experiment for the medical domain using speech from Indian doctors and medical students. The experiment used 855 clips for tuning and evaluated on 2,764 clips that included speakers different from those in the tuning data. As a result, the WER of 22.44% obtained with the best-performing prompt fell to 12.17% after additional training, a relative error reduction of 45.8%.
There is, however, another side to domain-specific gains. On LibriSpeech clean, a general speech recognition dataset, the WER worsened from 1.54% before additional training to 1.69%. On some evaluation items, such as pharmaceutical brand names, other models specialized for the medical domain came out ahead. This shows that while adapting to a specific domain can bring large improvements, performance on general speech may decline. These results are for Indian medical speech, and it has not been demonstrated that similar improvements would be obtained in Japanese.
If you are considering adopting Nemotron 3.5 Transcribe, first use the recordings you actually handle to test how effective terminology specification is. For multi-speaker recognition, compare performance including overlapping speech and microphone placement, and measure the throughput obtained with your GPU and the public implementation. If you perform additional training, check whether the improvement holds on speech from speakers not used in training, and whether accuracy on general recordings has degraded.
Nemotron 3.5 Transcribe is distinguished by going beyond a simple race for transcription accuracy, producing records that reflect technical terms, notation rules, and speaker information. Whether that advantage can be realized in actual operation, however, will depend on recording conditions, target languages, and computing resources. Going forward, independent comparative evaluations that include Japanese, and testing under conditions close to real-world use with the public implementation, will be important inputs for adoption decisions.