NVIDIA Nemotron 3.5 Transcribe 与 Granary v2：可提示、可适配的智能多语言转录

即使逐字记录了所有口头内容，用户或应用仍可能不清楚是谁说了什么、该如何翻译，或如何规范地格式化特定领域的术语。多语言通话、快节奏会议和医疗口述各自需要不同的输出：翻译、说话人标签，或拼写正确的医学术语。A transcript can capture every spoken word and still leave users or applications wondering who said what, how to translate it, or how to format domain-specific jargon. Multilingual calls, fast-paced meetings, and medical dictation each need different output: a translation, speaker labels, or correctly spelled medical terms.

NVIDIA Nemotron 3.5 Transcribe 是一款开放权重、可通过提示词引导的模型，能够将录制的语音转换为实用文本，现已在 Hugging Face 和 ModelScope 上推出。它将转录、说话人归属（说话人分离）、翻译和音频问答整合到一个模型检查点中：支持转录 40 种语言，将其中 31 种语言与英语互译，并能识别 13 种语言中的说话人。它可处理长达一小时的录音，并标注最多八位说话人。开发者可以用自然语言引导输出、选择音频的处理位置，还能根据领域词汇和口音调整识别效果。NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that turns recorded speech into useful text, now available on Hugging Face and ModelScope. It brings transcription, speaker attribution (diarization), translation, and audio question answering into one checkpoint: it transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour long and labels up to eight speakers. Developers can guide output with natural language, choose where their audio is processed, and adapt recognition to domain vocabulary and accents.

在 Hugging Face 的 Open ASR Leaderboard 英语榜单上，Transcribe 在开放权重模型中名列第一，并位于 Artificial Analysis 的准确率—活跃参数量前沿。在说话人归属方面，它在 David AI 的 21 种语言说话人分离排行榜上取得最低错误率，并在 VoiceArena 的 14 个系统对比中排名第一。它可在单张 NVIDIA GPU 上运行，适用范围从 DGX Spark 到 Blackwell 数据中心系统。Transcribe is first among open-weight models on the Hugging Face Open ASR Leaderboard for English and sits on Artificial Analysis’s accuracy–active-parameter frontier. For speaker attribution, it posts the lowest error rate on David AI’s 21-language diarization leaderboard and ranks first in VoiceArena’s 14-system comparison. It runs on a single NVIDIA GPU, from DGX Spark to Blackwell data-center systems.

Granary v2 的多语言数据（可在 Hugging Face 和 ModelScope 上获取）与可复用的数据整理流程，结合 NeMo Speech 微调工具，将提示级控制扩展到了训练阶段：开发者可以定义所需的书面形式，并使模型适应其应用中遇到的语音。在一项医疗案例研究中，与效果最佳的提示相比，微调使模型在留出说话人上的词错误率（WER）相对降低了 45.8%。Granary v2’s multilingual data (available on Hugging Face and ModelScope) and reusable curation pipeline, together with NeMo Speech fine-tuning tools, extend prompt-level control to training: developers can define the written form they need and adapt the model to the speech their applications encounter. In a medical case study, fine-tuning cut word error rate (WER) on held-out speakers by 45.8% relative to the best prompt.

同一个检查点可以根据指令，将一段会议录音转换成不同的输出：The same checkpoint can turn one meeting recording into different outputs, depending on the instruction:

• 转录这段录音，并标注每位说话者。
  “Transcribe this recording and label each speaker.”
• 将这段音频翻译成英语。
  “Translate this audio into English.”
• 这次会议作出了哪些决定？
  “What decisions were made in this meeting?”

在下面的视频中，看看这个模型如何应对 1940 年疯狂喜剧《女友礼拜五》中的经典场景——这部电影以语速飞快、对话频频交叠而闻名。Experience the model in action in the video below, where it tackles a classic scene from the 1940 screwball comedy His Girl Friday—a film famous for its rapid-fire, heavily overlapping dialogue.

Nemotron 3.5 Transcribe 为《女友礼拜五》（1940）中的一个场景生成字幕，并标注说话者，包括重叠对话。 Nemotron 3.5 Transcribe captions a scene from His Girl Friday (1940) and labels speakers, including overlapping dialogue.

对于不熟悉的术语，先结合上下文判断，再进行训练。例如，在医疗保健场景中：For unfamiliar terms, try context before training. In a healthcare setting, for example:

用原始语言转录这段医学讨论。可能涉及的术语：HbA1c、二甲双胍。只有在音频内容支持的情况下才使用这些术语。保留标准医学拼写，并以常规书写形式写出数字。Transcribe this medical discussion in the original language. Possible terms: HbA1c, metformin. Use these terms only when supported by the audio. Preserve standard medical spelling and write numbers in conventional written form.

模型卡中提供了推荐预设，包括用于音频问题推理的预设，还包含完整的提示指南和按任务与语言划分的矩阵：不同任务的语言支持各不相同，而且音频问答是用英语训练的。该模型还支持词级时间戳。与任何语音模型一样，请对照录音抽查人名、说话人变化和回答。The model card has recommended presets, including reasoning for audio questions, plus the full prompt guide and a task-by-language matrix: language support differs by task, and audio question answering is trained in English. Word-level timestamps are also supported. As with any speech model, spot-check names, speaker changes, and answers against the recording.

从实时课堂到长久留存的课程。Agora 的实时语音和视频平台为应用中的实时对话提供支持。在一场课堂演示中，Agora 使用了词级时间戳：Nemotron 3.5 Transcribe 将录制的课程转换为转录文本，并与回放并排显示，让学生了解讲了什么、何时讲的。Agora 的博客详细介绍了他们的构建过程。From live classes to lasting lessons. Agora, whose real-time voice and video platform powers live conversations in apps, used word-level timestamps in a classroom demo: Nemotron 3.5 Transcribe turns a recorded class into a transcript shown alongside the replay, so students can see what was said and when. Agora’s blog walks through how they built it.

在开放权重模型中，Transcribe 在识别说了什么方面领先；在判断是谁说的方面，它更是毫无悬念地登顶了两项说话人归属基准测试。Among open-weight models, Transcribe leads on what was said; on who said it, it tops two speaker-attribution benchmarks outright.

Hugging Face 开放式 ASR 排行榜。该排行榜在公开测试集上以词错误率（WER）衡量转录准确度，分为三个赛道：Hugging Face Open ASR Leaderboard. The leaderboard measures transcription accuracy as word error rate (WER) on public test sets, in three tracks:

1. 英语：以 5.33% 的平均 WER 位居开放权重模型之首（图 1）。
   English: first among open-weight models at 5.33% average WER (Figure 1).
2. 多语言：在模型卡的 Open ASR 对比中，七种语言的平均 WER 为 4.62%，是受评估开放权重模型中最低的。
   Multilingual: 4.62% mean WER across seven languages, the lowest among evaluated open-weight models in the model card’s Open ASR comparison.
3. 长文本：平均 WER 为 10.98%，也是该比较中最低的。
   Long-form: 10.98% mean WER, also the lowest in that comparison.

图 1. Nemotron 3.5 Transcribe 在 Hugging Face Open ASR Leaderboard（英语）上的开源权重模型中排名第一，平均词错误率（WER）为 5.33%，不包括 API 模型。WER 越低，表示转录错误越少。Figure 1. Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard (English) at 5.33% average WER, with API models excluded. Lower WER means fewer transcription errors.

Artificial Analysis 语音转文本排行榜。Artificial Analysis 根据 AA-WER（其词错误率指数）对语音转文本模型进行独立排名。Nemotron 3.5 Transcribe 在开放权重模型中并列第一，AA-WER 为 2.8%；而与它并列的模型有 24B 参数，它的活跃参数约为 4B。在 AA-AgentTalk（一个留出语音代理语音集，占 AA-WER 指数的 50%）上，Transcribe 的 WER 为 1.8%，而表现次优的开放权重模型为 2.3%。Transcribe 将语音编码器与混合式 Mamba–Transformer 混合专家语言模型结合，因此每个 token 仅激活其 30B 参数中的约 4B。这使它跻身 Artificial Analysis 开放权重模型在准确率与活跃参数之间的帕累托前沿。在 AA 的全模型图表中，它的排名高于多个广泛使用的专有语音 API。Artificial Analysis Speech-to-Text Leaderboard. Artificial Analysis independently ranks speech-to-text models by AA-WER, its word error rate index. Nemotron 3.5 Transcribe ties for first among open-weight models at 2.8% AA-WER, with about 4B active parameters versus 24B for the model it ties. On AA-AgentTalk, the held-out voice-agent speech set that makes up 50% of the AA-WER index, Transcribe scores 1.8% WER, versus 2.3% for the next-best open-weight model. Transcribe pairs its speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model, so only about 4B of its 30B parameters are active per token. That puts it on Artificial Analysis’s open-weight Pareto frontier of accuracy versus active parameters. On AA’s all-model chart, it ranks ahead of several widely used proprietary speech APIs.

图 2. 仅使用约六分之一的活跃参数，实现领先的 AA 准确率：Nemotron 3.5 Transcribe 的 AA-WER 为 2.80%，活跃参数约为 4B。Figure 2. Leading AA accuracy at about one-sixth the active parameters: Nemotron 3.5 Transcribe achieves 2.80% AA-WER with approximately 4B active parameters.

跨语言和重叠语音中的说话人归属。Sortformer 的说话人提示会与声学特征一同传入语言模型，因此 Transcribe 可以在生成过程中为每个轮次标注说话人，而不必完全依赖后续步骤为词语添加说话人标签。Speaker attribution across languages and overlap. Sortformer speaker cues reach the language model with the acoustic features, so Transcribe can tag each turn during generation instead of relying solely on a later step to attach speakers to words.

David AI DAI-ASR-I18N 排行榜。在 David AI 的 DAI-ASR-I18N 转录与说话人归属综合赛道中，Transcribe 在受评估的开放权重模型中排名第一，综合排名第二。在该基准测试的说话人分离赛道中，系统需要评估 21 种语言中每个人的发言时间；在测试的 14 个系统中，Transcribe 的说话人分离错误率最低，略优于 NVIDIA 专用的 Nemotron 3 Diarization 模型。David AI DAI-ASR-I18N Leaderboard. On David AI’s DAI-ASR-I18N combined transcription-and-speaker-attribution track, Transcribe ranks first among evaluated open-weight models and second overall. On the benchmark’s diarization track, which scores who spoke when across 21 languages, it has the lowest diarization error rate of the 14 systems tested, slightly ahead of NVIDIA’s dedicated Nemotron 3 Diarization model.

图 3. NVIDIA 模型占据 David AI 多语言说话人分离排行榜的前两名：Nemotron 3.5 Transcribe 的总体说话人分离错误率（DER）最低，在 21 种语言中的 DER 为 18.26%；其次是 Nemotron 3 Diarization，为 18.38%。DER 用于衡量说话人时间段错误；数值越低越好。Figure 3. NVIDIA models hold the top two spots on David AI’s multilingual diarization leaderboard: Nemotron 3.5 Transcribe posts the lowest overall diarization error rate (DER), 18.26% across 21 languages, followed by Nemotron 3 Diarization at 18.38%. DER measures speaker-time errors; lower is better.

VoiceArena 说话人分离基准测试。在 VoiceArena 的英语全会话基准测试中，Transcribe 的拼接最小排列词错误率（cpWER）在 14 个系统中最低，并且在面对面和在线两种场景下均排名第一。它漏掉的词也最少：删除词占参考词的 8.32%，而中位数约为 31%。VoiceArena Diarization-Bench. On VoiceArena’s English full-session benchmark, Transcribe posts the lowest concatenated minimum-permutation WER (cpWER) of 14 systems and ranks first in both in-person and online settings. It also drops the fewest words: deletions are 8.32% of reference words, versus a median of about 31%.

图 4. 在 VoiceArena 的 14 个系统对比中，NVIDIA 模型包揽前两名：Nemotron 3.5 Transcribe 以 32.4% 的 cpWER 领先，其次是搭配 Nemotron 3 Diarization 的 NVIDIA Multitalker Parakeet，cpWER 为 34.2%。Diarization-Bench v1 涵盖英语完整会话、重叠语音以及 2–8 位说话人。cpWER 在匹配说话人标签后统计词语错误数；数值越低越好。Figure 4. NVIDIA models hold the top two spots in VoiceArena’s 14-system comparison: Nemotron 3.5 Transcribe leads at 32.4% cpWER, followed by NVIDIA Multitalker Parakeet with Nemotron 3 Diarization at 34.2%. Diarization-Bench v1 covers English full sessions, overlapping speech and 2–8 speakers. cpWER counts word errors after matching speaker labels; lower is better.

更多评测。Ocular AI 的 Converse-STT 评测将 Transcribe 列为排名最高的开放模型。Josh Talks 的 Voice of India 排行榜展示了 Transcribe 在印度各语言中的表现。More evaluations. Ocular AI’s Converse-STT evaluation ranks Transcribe the top open model. Josh Talks’ Voice of India leaderboard reports Transcribe’s results across Indian languages.

从边缘端到数据中心均可运行。Transcribe 可在单块 NVIDIA GPU 上运行，涵盖 Jetson AGX Thor、DGX Spark，直至 Blackwell 数据中心 GPU。在一块采用 NVFP4 精度的 NVIDIA RTX PRO 6000 Blackwell Server Edition GPU 上，它处理 100 小时音频仅需不到三分钟；该数据基于高并发条件下的 30 秒音频片段测得。Runs from edge to data center. Transcribe runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs. On one NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with NVFP4 precision, it processes 100 hours of audio in under three minutes, measured on 30-second clips at high concurrency.

今天可以使用 vLLM 部署该模型检查点，或通过 NeMo-Speech.cpp 在本地运行。Transcribe 专为录制音频设计，而非实时流式传输。Serve the checkpoint with vLLM today or run it locally with NeMo-Speech.cpp. Transcribe is built for recorded audio, not live streaming.

特定领域的录音仍可能需要适配领域词汇和转录规范。Granary v2 和 NeMo Speech 微调正是为了解决这一问题，而这一切从支撑 Transcribe 的数据着手。Specialized recordings can still call for adaptation to domain vocabulary and transcript conventions. Granary v2 and NeMo Speech fine-tuning address that, starting with the data behind Transcribe.

可提示语音模型需要格式目标明确无歧义的训练数据。Granary v2 是 Hugging Face 和 ModelScope 上 Nemotron 语音数据集的一部分。该数据集使用 NVIDIA NeMo Curator 构建，其可复用流水线处理了多语言语音数据，供 Nemotron 3.5 Transcribe 训练使用，并生成对齐的转录文本变体、任务标签以及可检查的质量元数据。公开清单涵盖 48 种语言、来自 16 个源语料库的 463,062 小时数据；Granary v1 则涵盖 25 种语言。这些数字描述的是该数据集版本的发布规模，并不代表 Transcribe 的完整训练数据组合或其采样权重。A promptable speech model requires training data with unambiguous formatting targets. Granary v2 is part of the Nemotron speech collection on Hugging Face and ModelScope. Built with NVIDIA NeMo Curator, its reusable pipeline prepared multilingual speech used in Nemotron 3.5 Transcribe’s training, producing aligned transcript variants, task labels, and inspectable quality metadata. The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. These figures describe the dataset release, not Transcribe’s full training mixture or its sampling weights.

Granary v2 会为每段录音生成三种转录文本（表 1），即开发者可通过提示词从 Transcribe 请求的逐字稿和清理稿两种风格。除转录外，该流水线还会生成简短的音频描述、对借词进行语码转换还原、基于语音的问答对、用于上下文 ASR 偏置的术语，以及带有自动质量评分的翻译。Granary v2 generates three transcript forms for each recording (Table 1), the same verbatim and clean styles developers can request from Transcribe with a prompt. Beyond transcription, the pipeline produces short audio captions, code-switching restoration of loanwords, speech-grounded Q&A pairs, contextual ASR biasing terms, and translations with automatic quality scores.

表 1. 同一话语的 Granary v2 训练目标。Table 1. Granary v2 training targets for the same utterance.

该流水线由四个可复用的阶段组成：The pipeline is organized into four reusable stages:

• 1. 音频准备（元数据）：将原始音频或长音频切分为多个片段，在整个录音中跟踪说话人，并标注声学事件和语言。Granary v2 的音频源已完成切分，因此本次发布跳过了这一阶段。
  1. Audio Preparation (Metadata): Segments raw or long-form audio, tracks speakers across the full recording, and tags acoustic events and language. Granary v2’s sources were already segmented, so this release skipped this stage.
• 2. 转录（ASR）：为每种语言指定主 ASR 模型和备用 ASR 模型；如果主模型未通过幻觉检测，则切换到备用模型。人工参考转录文本有助于保留可信的拼写，同时补充口语形式和言语不流畅现象。
  2. Transcription (ASR): Routes each language to primary and recovery ASR models, falling back when the primary fails a hallucination check. Human reference transcripts help preserve trusted spellings while adding spoken forms and disfluencies.
• 3. 任务标签生成（文本）：生成转录文本的多个变体，然后检查字符率、幻觉模式，以及与模型或人工参考的一致性。只有通过审核的记录才会获得上下文 ASR 术语和翻译。
  3. Task Label Generation (Text): Derives transcript variants, then checks character rate, hallucination patterns, and model or human-reference agreement. Only accepted records receive contextual-ASR terms and translations.
• 4. 高保真（HiFi）：过滤掉存在噪声、截断或包含多个说话者的片段，保留较小规模的高质量录音集合，用于 ASR 以外的用途。对于 VoxPopuli 数据集，该筛选将 35.4 万小时的音频缩减至覆盖 22 种欧洲语言的 11.1 万小时。
  4. High Fidelity (HiFi): Filters out segments that are noisy, clipped, or contain multiple speakers, keeping a smaller set of high-quality recordings for uses beyond ASR. For the VoxPopuli dataset, it reduced 354k hours to 111k hours across 22 European languages.

图 6. Granary v2 数据整理流程：追踪一个样本从原始音频开始，依次经过元数据、ASR 和文本阶段，最终成为训练样本。Figure 6. Granary v2 data curation pipeline, following one sample from raw audio through the Metadata, ASR, and Text stages to a training example.

开发者可以使用 Ray 和 NVIDIA Dynamo 后端，在自己的录音或获准使用的公开语料上运行 NeMo Curator 流水线，也可以基于 Granary v2 公共清单构建。清单包含三种转录形式（tn_raw、itn_raw 和 itn_clean）、上述标签、实体提示、派生的指令对、来源信息，以及每条记录的质量标记。请根据各语料原有的许可单独下载源音频，并遵守该版本发布的条款。Developers can run the NeMo Curator pipeline with Ray and NVIDIA Dynamo backends on their own recordings or authorized public corpora, or build on the Granary v2 public manifests. The manifests include the three transcript forms as tn_raw, itn_raw, and itn_clean, the labels described above, entity hints, derived instruction pairs, provenance, and a per-record quality flag. Download source audio separately under each corpus’s original license, and follow the release’s published terms.

从一个明确指定语言、文字体系和转录格式的开发集提示词开始。在下面的案例研究中，在提示词中注明英语和拉丁字母，将全基准测试的 WER 从 30.8% 降至 23.8%。对于仍然存在的领域错误，NeMo Speech 微调工作流会将有代表性的语音与明确的书面目标配对。Start with a development-set prompt that specifies language, script, and transcript format. In the case study below, naming English and the Latin alphabet in the prompt cut full-benchmark WER from 30.8% to 23.8%. For domain errors that remain, the NeMo Speech fine-tuning workflow pairs representative speech with explicit written targets.

NeMo Speech 医疗 ASR 案例研究展示了这一进展。研究使用 Eka Care 的开放基准数据集，该数据集包含印度医生和医学生录制的音频，共 3,619 段，且未指定训练集划分。研究利用生成的医学文本和公开药品目录构建适配数据，随后结合约 153 小时的合成语音、29 小时来自 Common Voice 的带口音英语语音，以及 63 小时来自 MultiMed 的医疗录音。The NeMo Speech medical ASR case study shows this progression on Eka Care’s open benchmark of recordings from Indian doctors and medical students, which has 3,619 clips and no designated training split. The study builds adaptation data from generated medical text and public medicine catalogs, then combines approximately 153 hours of synthetic speech, 29 hours of accented English from Common Voice, and 63 hours of medical recordings from MultiMed.

图 7. 在 Eka Care 印度医疗语音基准的完整测试集上，更好的提示词将 WER 从 30.8% 降至 23.8%，微调后进一步降至 11.8%，比开箱即用的模型减少约 62% 的转录错误。完整测试集包含用于选择提示词和检查点的 855 段调优音频；下方的留出集结果才是更纯粹的测试结果。Figure 7. On the full test set of Eka Care’s Indian medical speech benchmark, a better prompt cuts WER from 30.8% to 23.8%, and fine-tuning brings it to 11.8%, about 62% fewer transcription errors than the model out of the box. The full set includes the 855 tuning clips used to choose the prompt and checkpoints; the held-out results below are the cleaner test.

合成生成将书面目标文本与口语呈现形式区分开来：“HbA1c”是标签，而“ H B A one C”则被发送给语音合成器。使用多种 TTS 语音可以丰富音频；真实录音则增添了不同的口音和录制条件。训练文本会经过筛选，以排除与基准参考文本完全匹配或共享八词序列的内容。Synthetic generation separates the written target from its spoken rendering: “HbA1c” is the label, while “H B A one C” is sent to the synthesizer. Multiple TTS voices broaden the audio; real recordings add accents and recording conditions. Training text is filtered for exact matches and shared eight-word sequences with benchmark references.

与 Granary v2 的明确目标一样，训练示例会变换数字、单位间距、公司名称全称以及品牌名称中的连字符，并通过指令指定所需的规范。有些提示中包含术语列表，其中夹杂着容易混淆的干扰项，让模型学会仅在音频内容支持时才采用相应拼写。最终系统使用一个固定的规范提示，并运行一次模型，不使用检索或对转录文本进行后处理。As with Granary v2’s explicit targets, training examples vary digits, unit spacing, company-name expansion, and brand-name hyphens, with instructions specifying the required convention. Some prompts include term lists with near-miss distractors, so the model learns to use a spelling only when the audio supports it. The final system uses one fixed convention prompt and one model pass, without retrieval or transcript post-processing.

该方案训练语音编码器和投影层，并使用 LoRA 适配语言模型。两次运行使用不同的数据混合；合并适配器后，将选定的检查点进行均匀平均，得到一个最终模型。检查点的选择依据是开发集解码结果，而非仅看训练损失。从文本生成到评估的完整适配过程，在一块 NVIDIA GB300 GPU 上耗时约 14 GPU 小时。The recipe trains the speech encoder and projection and uses LoRA to adapt the language model. Two runs use different data mixtures; after adapter merging, selected checkpoints are uniformly averaged into one final model. Checkpoint selection uses decoded development-set results, rather than training loss alone. The full adaptation, from text generation through evaluation, took about 14 GPU-hours on one NVIDIA GB300.

在留出说话人上测得的领域增益 Domain gains, measured on held-out speakers

该研究按说话人划分基准数据：855 段调优音频用于选择提示词和检查点，2,764 段留出音频则由每个最终系统各解码一次。使用 Eka Care 的 KARMA 评分流程，微调将留出集 WER 从最佳提示词基线的 22.44% 降至 12.17%，相对降幅为 45.8%，这是性能提升能够泛化到新说话人的最有力证据。关键词 WER 从 9.2% 降至 6.5%，语义 WER 从 10.0% 降至 7.6%。在相同的留出说话人和评分器上，适配后的模型 WER 为 12.17%，优于开放的医学专用模型（13.79%）和广泛使用的通用 ASR 模型（14.92%）。The study splits the benchmark by speaker into 855 tuning clips, used to choose the prompt and checkpoints, and 2,764 held-out clips, decoded once per final system. Using Eka Care’s KARMA scoring pipeline, fine-tuning cuts held-out WER from 22.44% for the best-prompt baseline to 12.17%, a 45.8% relative reduction and the clearest evidence that the gains generalize to new speakers. Keyword WER falls from 9.2% to 6.5%, and semantic WER from 10.0% to 7.6%. On the same held-out speakers and scorer, the adapted model’s 12.17% WER beats an open medical-specialist model (13.79%) and a widely used general-purpose ASR model (14.92%).

在通用英语方面，LibriSpeech test-clean 的 WER 略有上升，从 1.54% 升至 1.69%。医疗专业模型在品牌药物和诊断项目的精确匹配上仍然领先；该研究指出，下一步应使用临床医生说出品牌名称的真实录音。对于你自己的领域，至少保留 100 段真实且带标签的音频片段用于检查点选择，并在评估通用语音能力保留情况的同时，评估应用所依赖的实体类型。On general English, LibriSpeech test-clean WER rises slightly, from 1.54% to 1.69%. The medical specialist still leads on exact matches for brand-name drugs and diagnostics; the study points to real recordings of clinicians saying brand names as the next step. For your own domain, keep at least 100 real, labeled clips for checkpoint selection, and evaluate the entity types your application depends on alongside general-speech retention.

该案例研究包含用于复现每个步骤的命令。微调教程涵盖训练、导出和评估的方案、资源需求与工具，包括新语言和语码转换 ASR。配套的智能体技能可让编程智能体运行整个工作流，从基线选择和数据检查，到训练验证、适配器合并与评估。The case study includes commands to reproduce each step. The fine-tuning tutorials cover recipes, resource requirements, and tools for training, export, and evaluation, including new-language and code-switching ASR. The accompanying agent skill lets a coding agent run the workflow, from baseline selection and data checks through training validation, adapter merging, and evaluation.

ModelScope 还支持开发者通过其 ms-swift 框架微调 Nemotron 3.5 Transcribe，以用于专门的多模态应用。ModelScope also enables developers to fine-tune Nemotron 3.5 Transcribe through its ms-swift framework for specialized multimodal applications.

在 Hugging Face 演示中或 ModelScope Studios 上试用 Nemotron 3.5 Transcribe，然后从 Hugging Face 或 ModelScope 下载检查点，并按照 vLLM 部署指南部署并提供服务。Try Nemotron 3.5 Transcribe in the Hugging Face demo, or on ModelScope Studios, then download the checkpoint from Hugging Face or ModelScope and serve it with the vLLM deployment guide.

如需跳过自托管，请使用 Baseten、DeepInfra 或 Simplismart 提供的托管端点。To skip self-hosting, use a hosted endpoint from Baseten, DeepInfra, or Simplismart.

要将其适配到你的领域，可以从 Hugging Face 或 ModelScope 上的 Granary v2 清单和医疗 ASR 案例研究入手，也可以让编码代理借助微调代理技能来执行整个工作流。To adapt it to your domain, start with the Granary v2 manifests on Hugging Face or ModelScope and the medical ASR case study, or let a coding agent drive the workflow with the fine-tuning agent skill.

模型与数据集Models & Datasets

• Model Checkpoints: NVIDIA Nemotron 3.5 Transcribe on Hugging Face, ModelScope | Interactive Demo on Hugging Face, ModelScope
• Hosted Endpoints: Baseten, DeepInfra, Simplismart \
• 数据集清单：Hugging Face、ModelScope 上的 Granary v2 清单
  Dataset Manifests: Granary v2 Manifests on Hugging Face, ModelScope
• 高保真子集：Hugging Face、ModelScope 上的清单 | Hugging Face 上的空间
  High Fidelity Subset: Manifests on Hugging Face, ModelScope | Space on Hugging Face

代码库与教程：NVIDIA NeMo Speech GitHub | NeMo Curator Granary v2 流程 | vLLM 部署指南 | 语音微调指南 | 微调智能体技能 *合作伙伴博客：Agora | Inflection AI | VoiceArena | Simplismart * 本地 AI：NVIDIA RTX PC、笔记本电脑和 DGX Spark 上的本地 AI 有哪些新动态Repositories & Tutorials: NVIDIA NeMo Speech GitHub | NeMo Curator Granary v2 Pipeline | vLLM Deployment Guide | Speech Fine-Tuning Guide | Fine-Tuning Agent Skill
*Partner Blogs: Agora | Inflection AI | VoiceArena | Simplismart *
Local AI: What’s new in local AI on NVIDIA RTX PCs, laptops, and DGX Spark