NVIDIA Releases Nemotron 3 Diarization with 8-Speaker Support and Streaming Latency Down to 0.32 Seconds

NVIDIA Releases Nemotron 3 Diarization with 8-Speaker Support and Streaming Latency Down to 0.32 Seconds NVIDIA has introduced Nemotron 3 Diarization, an open-weight, 100M-parameter model capable of distinguishing up to eight speakers in real-time and recorded audio. The model tops VoiceArena's Diarization-Bench leaderboard with a 14.72% Diarization Error Rate (DER) and provides configurable latency modes down to 0.32 seconds. NVIDIA Nemotron 3 Diarization features a 31-layer Transformer encoder with RoPE, expanding speaker support from earlier 4-speaker models to eight speakers. The model ranked #1 on VoiceArena's Diarization-Bench leaderboard with a 14.72% DER, marking a ~24% relative error reduction over the runner-up. Streaming inference supports four recommended latency configurations ranging from an offline-style 30.4 seconds down to an ultra-low latency of 0.32 seconds. Training incorporated public data and licensed multilingual mixtures from David AI across 21 languages, lowering compound DER from 11.19% to 10.42%. Model outputs anonymous speaker channel probabilities rather than real-world identities, requiring downstream components for verification and transcription alignment. Architecture and Streaming Mechanics Nemotron 3 Diarization accepts 16 kHz, single-channel audio, converting it to Mel-spectrogram features at 10 ms steps before stacking them eightfold into 80 ms frames. These feed into a 31-layer Transformer encoder equipped with rotary positional embeddings (RoPE), followed by a Conv1D layer that outputs a [T, 8] tensor of active speaker probabilities. Following the Sortformer approach, it orders speakers by arrival time to avoid permutation matching across chunks. For streaming inference, context is maintained through an Arrival-Order Speaker Cache (AOSC), a FIFO buffer of recent frames, and chunk right-context. Benchmark Results and Throughput Across 139 English conversations in VoiceArena's Diarization-Bench, the model achieved first place across 12 systems and 17 configurations with a 14.72% DER under a 0 ms collar with overlapping speech scored. In multi-dataset benchmarks against the streaming Sortformer v2.1 baseline at 1.04-second latency, Nemotron 3 Diarization yielded an unweighted mean relative DER reduction of 41.0% across eight conditions. Using BF16 and torch.compile() on an NVIDIA RTX PRO 5000, it achieved throughputs of 15,113× RTFx at 30.4s latency and 865× RTFx at 1.04s latency at batch size 32. Deployment and Practical Limitations The model is governed by the OpenMDW License Agreement v1.1 and supports Linux setups with Ampere, Hopper, or Blackwell GPUs, as well as deployment through Baseten, Digital Ocean, and Argmax Pro SDK 3. The outputs assign anonymous channel labels rather than named identities and cap capacity at eight speakers; exceeding this capacity leads to missed speech or channel confusion. Furthermore, heuristics like the midpoint rule do not separate overlapping speech or establish which speaker uttered an ASR word, and performance may degrade under heavy reverberation, severe noise, far-field recording, or domain shift.

NVIDIA 推出开源 Nemotron 3 Diarization 说话人日志模型，支持实时多说话人识别 NVIDIA 发布了开源权重的 Nemotron 3 Diarization 说话人日志模型，参数量为 100M，在 VoiceArena Diarization-Bench 榜单中以 14.72% 的日记化错误率（DER）位居第一。该模型沿用 Sortformer 架构，支持最多 8 位说话人、重叠语音检测与灵活的分块流式推理。 基于 100M 参数的 31 层 Transformer 架构，采用按说话人出现时间排序的机制，输出多达 8 个说话人通道的激活概率。 在 VoiceArena 评测中 DER 为 14.72%，较次席系统相对降低约 24%，在 1.04 秒低延迟配置下多数据集表现显著优于基线模型。 提供从 30.4 秒到 0.32 秒等推荐输入缓冲延迟配置，在 RTX PRO 5000 上使用 NeMo PyTorch 后端实现了极高的批处理 RTFx 吞吐量。 模型输出为匿名通道而非实体身份，最多支持 8 人，在长音频、高噪声或混响等复杂声学环境下性能可能出现退化。 架构设计与流式机制 Nemotron 3 Diarization 接收 16 kHz 单声道音频，通过 10 ms 步长提取梅尔频谱，以 8 倍堆叠形成 80 ms 帧输入带 RoPE 的 31 层 Transformer 编码器，最终由 Conv1D 恢复分辨率并输出 [T, 8] 的激活概率张量。该模型沿用 Sortformer 方法，依据说话人出现时间排序以稳定标签并避免排列组合解算，同时利用到达顺序说话人缓存（AOSC）、先进先出队列及右上下文实现流式推理。训练数据涵盖多说话人真实对话以及来自 David AI 涵盖 21 种语言的合成混合数据，该数据的引入使复合 DER 绝对值降低了 0.77 个百分点。 基准评测与推理性能 在 VoiceArena 的 Diarization-Bench 初步评估中，Nemotron 3 Diarization 在 12 个系统和 17 种配置中排名第一，取得 14.72% 的 DER。在涵盖 DIHARD III、AliMeeting、AMI 等基准测试中，1.04 秒输入缓冲下的未加权平均 DER 较 Sortformer v2.1 基线相对降低 41.0%。在硬件性能方面，基于 NVIDIA RTX PRO 5000（BF16、torch.compile、批大小 32），模型在 30.4 秒离线延迟下达到 15,113× RTFx 和 12.73% DIHARD III DER，在 1.04 秒延迟下达到 865× RTFx 和 13.18% DER。Argmax Pro SDK 3 和 OpenBench 评测亦显示其在多数据集中取得了最低错误率。 配置选项与实际应用限制 模型提供了四种推荐延迟配置：30.4 秒离线式、1.04 秒低延迟、0.64 秒极低延迟以及 0.32 秒超低延迟，受 OpenMDW 许可协议 1.1 版约束。在应用落地方面，模型输出的仅为匿名通道，需要与元数据或声纹验证模型结合才能确定真实身份；模型上限为 8 位说话人，超出容量会导致漏检或分配错误。此外，简单的中点对齐启发式规则无法分离重叠语音，且长对话、噪声、强混响或远场拾音均可能加剧错误，因此在安全相关或关键业务流程中建议保留不确定性并进行针对性测试。

Turn overlapping conversations into speaker-aware data with one open-weight, 100M-parameter model - ranked #1 in VoiceArena's Diarization leaderboard with a 14.72% Diarization Error Rate (DER). Why speaker diarization matters Every conversation carries two layers of information: what was said and who said it . Speech recognition captures and transcribes the words. Speaker diarization classifies who spoke when , helping applications connect what was said to the right participant. Consider a transcript from a meeting, customer call, or podcast in which every sentence is correct but none is attributed to a speaker. You can read the words, but you cannot reliably tell who made a commitment, who raised an objection, or which participant interrupted. Search, summaries, action items, conversation analytics, and voice-agent memory all become less useful. Speaker diarization identifies the time intervals during which each speaker is active, including intervals where people talk over one another. Those speaker timestamps can then be combined with automatic speech recognition (ASR) to create a speaker-attributed transcript. NVIDIA Nemotron 3 Diarization is an open-weight, 100M-parameter model that ranks #1 on VoiceArena's Diarization-Bench leaderboard with a 14.72% Diarization Error Rate (DER). Supporting up to eight speakers across live and recorded conversations, it handles overlapping speech, chunked processing for flexible recording lengths, and customizable streaming latency. Figure 1. Nemotron 3 Diarization #1 on VoiceArena Diarization-Bench leaderboard. Earlier models like NVIDIA Streaming Sortformer established this approach for four-speaker diarization, including the streaming diar_streaming_sortformer_4spk-v2.1 checkpoint used as the baseline below. Nemotron 3 Diarization expands support to eight speakers and improves the accuracy and throughput measured in the following evaluations. How Nemotron 3 Diarization works One model for offline and streaming conversations Diarization systems must solve two related problems. First, they must detect speech and assign it to the correct speaker. Second, they must preserve that assignment throughout the conversation, even after silence, interruptions, or long gaps between a speaker's turns. Streaming makes the second problem harder. An offline model can inspect an entire recording at once. A streaming system receives only a small chunk of new audio and limited context. Without an effective memory mechanism, the speaker assigned to one output channel in the current chunk can be assigned to a different channel in the next. Nemotron 3 Diarization follows the Sortformer approach of ordering output speakers by when they first appear. The first new voice becomes the first speaker channel, the next new voice becomes the second, and so on. This arrival-time ordering makes the model's generic speaker labels stable and removes the need to solve a new speaker permutation for every chunk. Nemotron 3 Diarization was trained using public and licensed speech data, including multispeaker-annotated, real-world conversations licensed from David AI. Additional licensed David AI audio provided source material for large-scale simulated English and multilingual mixtures spanning 21 languages. Adding David AI data to our training decreased compound diarization error rate (DER) by 0.77 absolute points, from 11.19% to 10.42%, at both the offline-style and ultra-low-latency operating points. The model supports as many as eight speaker channels. These are anonymous labels, not real-world identities: the model can report that speaker_2 spoke from one timestamp to another, but it does not determine that speaker_2 is a particular person. Downstream applications can map these anonymous channel IDs to explicit speaker identities by pairing timestamps with meeting metadata, user profiles, or active speaker verification models. From audio to speaker activity The model accepts 16 kHz, single-channel audio and converts it into Mel-spectrogram features with a 10 ms frame step. It stacks those features by a factor of eight, producing 80 ms frames for a 31-layer Transformer encoder with rotary positional embeddings (RoPE). Figure 2. Nemotron 3 Diarization converts audio into arrival-ordered speaker-activity probabilities. AOSC and FIFO context support streaming inference. Above the Transformer, a Conv1D layer upsamples the predictions to the input-feature resolution. The default output is a [T, 8] floating-point tensor: T time steps by eight possible speaker channels. Each value is the probability that a speaker is active at that time. The default stride is 10 ms and can be configured to another multiple of 10 ms. This representation handles overlap naturally. If two people speak at the same time, two channels can be active in the same frame. Postprocessing converts these probabilities into generic speaker labels with start and end timestamps. During streaming inference, two forms of memory provide context: The Arrival-Order Speaker Cache (AOSC) retains useful information about speakers observed in earlier chunks, organized according to their arrival-ordered channels. A first-in, first-out (FIFO) queue supplies recent frame context before the current chunk. The input buffer also includes right context, which is audio immediately after the current chunk. More right context can help the model interpret speaker transitions, while less right context reduces the time it must wait before producing a result. Together, the current chunk, right context, FIFO queue, and speaker cache enable one model to operate at several latency points. Chunked inference removes a fixed model-imposed maximum audio duration. Performance can still degrade on unusually long recordings or audio with severe noise, reverberation, far-field capture, or domain shift. Diarization and Speaker Attributed ASR (Multi-speaker ASR) are different tasks Standalone diarization produces speaker activity and timestamps, not the words being spoken. ASR produces text but does not necessarily preserve speaker attribution. A speaker-attributed transcription pipeline combines both outputs: Figure 3. End-to-end speaker-attributed transcription pipeline combining audio diarization timestamps with automatic speech recognition (ASR) to map spoken words to specific speakers. That separation matters when designing a system. Diarization errors include missed speech, false speech detections, incorrect speaker assignments, and boundary errors. ASR errors affect the words. Applications should evaluate both components and the combined pipeline on their intended audio. Balance latency and accuracy The same model supports recommended input-buffer latencies of 30.4, 1.04, 0.64, and 0.32 seconds. Shorter buffers let the system respond sooner, while more context generally improves accuracy and throughput. These values measure the audio buffered before inference; computation, networking, ASR, and application processing add to end-to-end latency. Although the model can technically use an 80 ms input buffer, 0.32 seconds is the lowest recommended configuration. The configuration table in "Get started" shows how to select an operating point. Benchmark results: Ranked #1 in VoiceArena's initial Diarization-Bench In VoiceArena's initial Diarization-Bench results, NVIDIA Nemotron 3 Diarization ranked first among 12 systems and 17 total system configurations evaluated across 139 English-language conversations totaling approximately 22 hours. With overlapping speech scored, system-generated speech activity detection, and no boundary collar, Nemotron 3 Diarization achieved a 14.72% diarization error rate (DER), compared with 19.3% for the next-ranked system - a ~24% relative reduction. It also ranked first with 100 ms and 250 ms collars and across both in-person and online recordings. These initial results may change as Voice Arena completes its Version 1 evaluation and paired statistical analysis. Figure 4. Voice Arena Diarization-Bench v1 results (English, 0 ms collar, DER lower is better), showing NVIDIA Nemotron 3 Diarization ranking #1 with 14.72% DER. Measuring diarization accuracy The primary metric used to evaluate the model is diarization error rate (DER). It combines three kinds of error: Missed speech: a reference speaker was active, but the system detected no corresponding speech. False alarm: the system marked a speaker as active when the reference contained no corresponding speech. Speaker confusion: the system detected speech at the right time but assigned it to the wrong speaker. Figure 5. DER adds missed speech, false alarms, and speaker confusion, then divides by total reference speaker time. Overlapping reference speakers each contribute to the denominator. Benchmark settings can materially change DER, so the evaluation protocol is part of the result. The Nemotron 3 evaluation contains 901 condition-specific recordings spanning multilingual telephone speech, meetings, near-field and far-field microphones, multi-microphone capture, and difficult acoustic environments. Overlapping speech is scored in every evaluation. DIHARD III, AliMeeting, AMI, and NOTSOFAR1 use a zero-second collar, meaning no boundary tolerance is excluded from scoring. CALLHOME-Part2 uses a 0.25-second collar. Results were generated with the NeMo e2e_diarize_speech.py evaluation script. The comparison below uses diar_streaming_sortformer_4spk-v2.1, NVIDIA's previous four-speaker streaming Sortformer, as the baseline. It uses the final Nemotron-3-Diarization values. An average 40% relative DER reduction at 1.04-second latency Figure 6. Full-set DER for the model and previous NVIDIA baseline at 1.04-second input-buffer latency. Lower is better. At 1.04-second input-buffer latency, Nemotron 3 Diarization reduces DER on all eight listed evaluation conditions. The relative reductions range from 9.0% on CALLHOME-Part2 to 65.2% on NOTSOFAR1 MHM. The unweighted mean of the eight per-dataset relative reductions is 41.0% . In other words, this is an average of relative improvements across evaluation conditions; it is not a pooled DER computed by combining every recording into one score. The improvement is also consistent across operating points. At each latency shared by the two models (30.4, 1.04, and 0.32 seconds), the final model has lower full-set DER on every evaluated dataset. Figure 7. Full-set DER across input-buffer latency settings. The previous baseline does not have a 0.64-second configuration. Improvements increase in higher-speaker-count conditions Support for eight speakers enables meetings and group conversations with more than four participants. The benchmark advantage also widens in the higher-speaker-count subsets of DIHARD III, CALLHOME-Part2, and NOTSOFAR1. Figure 8. DER at 30.4-second input-buffer latency, broken out by speaker count. Shaded regions contain more than four speakers. The figure also preserves an important nuance: on the two-speaker CALLHOME subset, the final model records 5.98% DER compared with 5.68% for the previous baseline. Across the full CALLHOME-Part2 evaluation, however, DER improves from 10.32% to 9.10%, with larger gains in the higher-speaker-count subsets. DIHARD III groups recordings with five through nine speakers into one result. Nine speakers exceed Nemotron 3 Diarization's supported maximum of eight, so that aggregate includes audio outside the specified speaker-count limit. Accuracy and throughput A deployable diarization system must balance accuracy and throughput. The model card reports real-time factor speedup (RTFx), calculated as total audio duration divided by total processing time. Higher RTFx means the system processes more audio per unit of compute time. Figure 9. DIHARD III full-set DER versus batch-size-32 compiled throughput. Results use BF16 with the NeMo PyTorch backend on an NVIDIA RTX PRO 5000. At the 30.4-second configuration, Nemotron 3 Diarization reaches 15,113× RTFx at batch size 32 with torch.compile(), compared with 2,619× for the previous baseline, while lowering DIHARD III DER from 19.09% to 12.73%. At the 1.04-second configuration, it reaches 865× compared with 136×, while lowering DER from 19.60% to 13.18%. These results measure batched throughput on the disclosed test system. They should not be interpreted as single-stream, end-to-end application latency. Developers should benchmark the complete pipeline on their target hardware, including data movement, diarization, ASR, and downstream processing. See Nemotron 3 Diarization in action Explore the Nemotron Diarization live model demo to connect the model’s speaker timestamps with a conversation you can follow. The application offers synthetic conversations, an eight-speaker mode, live microphone input, and stress-test scenarios. We also added a live mic for a multlingual ASR models, so you can perform live streaming diarization with speakers of different languages. Start in Conversation with a preloaded topic, then follow the speaker activity and live transcript as the conversation plays. Listen for interruptions and compare the timing of the voices with the displayed speaker activity. For your own conversation, use Live Mic or the Multilingual Live Mic . Finally you can upload an audio file you have prerecorded in the Audio File tab. The following invented exchange illustrates the difference that speaker attribution makes; it is not a measured output from the demo: Without speaker attribution With speaker attribution "I’ll send the report." "Can you include the figures?" "Yes, by Friday." speaker_0: "I’ll send the report." speaker_1: "Can you include the figures?" speaker_0: "Yes, by Friday." With speaker labels, an application can associate the commitment with the same participant who answers the follow-up. The timeline adds information that plain text cannot show: two speakers can be active at once. A downstream summarizer can use the attributed transcript to extract action items while retaining the source speaker and timestamps. The prepared scenarios use synthetic audio. Some eight-speaker scenarios provide a short speaker-context primer before the audible conversation; they illustrate behavior with that context and are not an unprimed benchmark. Participant names and roles belong to the application’s scenario logic. Nemotron 3 Diarization supplies generic speaker channels and timestamps, not identity verification. Bring real-time, speaker-aware transcription on device Argmax has added support for NVIDIA Nemotron 3 Diarization in Argmax Pro SDK 3 , enabling real-time speaker attribution for conversations with up to eight speakers, and introduced a pre-diarized transcription API that separates speakers before speech recognition, designed to improve transcription in complex conversations with overlapping speech. We benchmarked 6 speaker diarization systems across 11 datasets representing diverse and challenging real-world conditions using OpenBench.Nemotron 3 Diarization achieved the lowest error rate (DER) across all 6 systems.It also scored the lowest across 5 of the 11 datasets despite its maximum speaker constraint of 8.This represents a huge leap in accuracy compared to its predecessor Sortformer v2.1 whose error rate was 60% higher. Atila Orhon Watch the demo to see Nemotron 3 Diarization running with Argmax Pro SDK 3. Figure 10 . Argmax demo video of the Nemotron 3 Diarization model Get started with NVIDIA NeMo Speech Below we will show a few quick start guides on how to combine diarization with ASR models, streaming ASR, diarization only and offline ASR with the NVIDIA Nemo Speech library . Install the dependencies Install the system packages and NVIDIA NeMo speech dependencies after setting up Python 3.12 or later, Cython, and a recent PyTorch version. apt-get update && apt-get install -y libsndfile1 ffmpeg uv pip install Cython packaging uv pip install 'nemo-toolkit[asr]' Implement streaming diarization with streaming ASR (multi-speaker ASR) in real-world scenarios To evaluate ASR and diarization (multi-speaker ASR) performance, refer to the Quick start guide to Diarization with streaming ASR . Run an offline diarization on a recording Use a 16 kHz mono recording containing two or more speakers, and replace /path/to/conversation.wav with its path. The following example loads the checkpoint and uses the recommended 30.4-second offline-style configuration. from nemo.collections.asr.models import SortformerEncLabelModel diar_model = SortformerEncLabelModel.from_pretrained( "nvidia/Nemotron-3-Diarization" ) diar_model. eval () # Configuration values are measured in 80 ms frames. diar_model.sortformer_modules.spkcache_len = 264 diar_model.sortformer_modules.fifo_len = 40 diar_model.sortformer_modules.chunk_len = 340 diar_model.sortformer_modules.chunk_right_context = 40 diar_model.sortformer_modules.spkcache_update_period = 300 diar_model._check_streaming_parameters() predicted_segments = diar_model.diarize( audio=[ "/path/to/conversation.wav" ], batch_size= 1 , ) for segment in predicted_segments[ 0 ]: print (segment) The diarize() method returns speaker-marked segments as strings in the form start_seconds end_seconds speaker_id. For example, the output structure can look like this (illustrative timestamps, not a measured model result): 0.400 2.100 speaker_0 1.800 3.250 speaker_1 3.600 4.700 speaker_0 Here, both speakers are active between 1.800 and 2.100 seconds. A speaker can appear in multiple segments, and the intervals need not be mutually exclusive. To obtain the underlying speaker-activity tensors as well, set include_tensor_outputs=True. The API accepts an audio path, a list of paths, NumPy arrays, or a line-delimited JSON manifest. When providing NumPy arrays, pass the correct integer sample_rate; the default file-audio sample rate is 16 kHz. Nemotron 3 Diarization supports 16 kHz, single-channel .wav, .flac, .opus, and .mp3 audio. It is designed for Linux systems with supported NVIDIA Ampere, Hopper, or Blackwell GPUs. Choose the latency-quality operating point Nemotron 3 Diarization exposes streaming parameters in units of 80 ms encoder frames. The recommended configurations cover offline-style processing through ultra-low-latency streaming. Configuration Input-buffer latency Speaker cache FIFO Chunk Right context Cache update period Offline style 30.4 s 264 40 340 40 300 Low latency 1.04 s 264 264 9 4 222 Very low latency 0.64 s 264 264 6 2 222 Ultra-low latency 0.32 s 264 264 3 1 222 The latency values in this table are input-buffer latency , calculated as:(CHUNK_LEN + RIGHT_CONTEXT) × 80 ms They do not include model computation, network transport, ASR, or application processing. Although the model can technically use an input buffer as short as 80 ms, 0.32 seconds is the lowest recommended configuration. Lowering latency generally reduces both accuracy and throughput, so developers should select the operating point against end-to-end product requirements rather than buffer duration alone. Use all five parameter values from the same row in the inference example, then call _check_streaming_parameters() before running diarization. The speaker cache retains earlier speaker context; FIFO controls the recent history; chunk and right context set the input-buffer latency; the cache update period controls how much FIFO context is used for a cache update. Combine diarization with Offline ASR Run ASR and diarization on the same recording and time base. For example, You can use Nemotron ASR 3.5 or Parakeet TDT 0.6B v3 that can return word timestamps. The following minimal offline example reuses predicted_segments from above and associates each word with the speaker active at its midpoint: from nemo.collections.asr.models import ASRModel asr_model = ASRModel.from_pretrained( model_name= "nvidia/parakeet-tdt-0.6b-v3" ) words = asr_model.transcribe( [ "/path/to/conversation.wav" ], timestamps= True )[ 0 ].timestamp[ "word" ] turns = [] for segment in predicted_segments[ 0 ]: start, end, speaker = segment.split() turns.append(( float (start), float (end), speaker)) def speaker_at ( midpoint ): active = sorted ({ speaker for start, end, speaker in turns if start <= midpoint < end }) if len (active) == 1 : return active[ 0 ] return "overlap/ambiguous" if active else "unassigned" for word in words: midpoint = (word[ "start" ] + word[ "end" ]) / 2 label = speaker_at(midpoint) print ( f" {word[ 'start' ]: .2 f} - {word[ 'end' ]: .2 f} {label} : {word[ 'word' ]} " ) This midpoint rule is a simple alignment heuristic. It marks simultaneous speaker activity as ambiguous and leaves words outside detected speech unassigned. It does not separate overlapping voices or determine which active speaker produced an ASR word. For a production transcript, evaluate word-boundary alignment, overlap handling, and both models’ errors on representative audio before combining adjacent words into speaker turns. Deployment considerations The model supports a maximum of eight speakers. If a recording contains more, speech can be missed or assigned to the wrong channel. Noise, severe reverberation, far-field recording, domain shift, and long conversations can also increase missed speech, false alarms, boundary errors, or speaker confusion. Downstream applications should preserve useful uncertainty instead of treating every speaker assignment as infallible. Evaluate the complete system on audio representative of the intended environment, especially before using speaker attribution in regulated, safety-relevant, or consequential workflows. Use of the model is governed by the OpenMDW License Agreement, version 1.1 . Resources and next steps Nemotron 3 Diarization on Hugging Face VoiceArena leaderboard Argmax Pro SDK 3 Announcement Argmaxinc Openbench NVIDIA NeMo Speech NeMo end-to-end diarization evaluation script Sortformer paper Streaming Sortformer paper Quick start guide to Diarization with ASR Interactive demo and sample application Parakeet ASR model and timestamp examples Build with NVIDIA ecosystem partners Deploy Nemotron 3 Diarization for production inference with Baseten Run Nemotron 3 Diarization at scale with Digital Ocean Build mobile and on-device speaker-attributed transcription with Nemotron 3 Diarization support in Argmax Pro SDK 3