NVIDIA NeMo Speech 3.0

NVIDIA NeMo Speech 3.0 Release Notes

NeMo Speech 3.0 is the first major release after the repo split and rename to NVIDIA-NeMo/Speech. The repo now focuses on ASR, TTS, audio processing, speaker tasks, and SpeechLM. Non-speech Framework, LLM, VLM, diffusion, export/deploy, evaluator components can now be found under separate repos under NVIDIA-NeMo organization.

While this release brings new major features, the central focus is addressing technical debt: removed 800k deprecated LOC, migration to uv package manager for cleaner installs, stronger model test coverage, reduced the number of dependencies, revamped documentation, lighter containers, and AGENTS.md + agentic skills.

Top-Level Changes and Highlights

Breaking Changes and Migration Notes

• LLM/VLM/diffusion/NLP/export/evaluator collections have been moved out to separate repos.
• Unsupported ASR/TTS models and old tutorials were removed. Use active examples under examples/ and docs under docs/source/.
• uv sync reproduces the supported NeMo Speech stack and may replace Python/PyTorch/CUDA inside .venv. To keep your own stack, install PyTorch first, then use uv pip or pip.
• SpeechLM2 compiled acceleration is optional. Use the supported Docker build for TE/FlashAttention/Mamba/grouped-GEMM/DeepEP performance paths.

SpeechLM2 and NeMo Automodel Backend

SpeechLM2 now supports training SpeechLM models with NeMo Automodel through SALMAutomodel, targeting both dense and MoE LLM backbones such as Nemotron 3.

• SALMAutomodel adds an Automodel-backed SALM path with native LoRA, deferred configure_model() initialization, shard-aware distributed loading, and Automodel-owned mesh creation (#15447).
• AutomodelParallelStrategy supports FSDP2, HSDP, TP, CP, and EP, letting SpeechLM training combine sharded data parallelism, tensor/model partitioning, long-context sequence sharding, and MoE expert routing (#15447, #15648, #15679, #15773).
• Optimized backend support now covers Transformer Engine, FlashAttention, Mamba/state-space kernels, grouped GEMM for MoE experts, and DeepEP for expert-parallel all-to-all communication (#15737, #15758, #15679).
• THD packed-sequence training reduces padding waste in variable-length speech batches and is the supported path for CP long-audio training (#15679).
• Long-audio support was expanded with encoder chunking, activation-checkpointing controls, CP-safe data handling, and fixes around dtype/checkpoint restoration (#15648, #15679, #15686, #15716).
• Export and serving improved with vLLM integration, HF/vLLM-ready checkpoint export, backbone-native chat templates, offline HF export, and buffered SALM inference (#15364, #15520, #15623, #15736).
• Data format support now includes ShareGPT conversations, indexed ShareGPT JSONL, and WebDataset-backed ShareGPT conversations (#15316, #15410).

Nemotron VoiceChat Training Modules

NeMo Speech 3.0 adds the train/eval modules behind Nemotron VoiceChat.

• Added Nemotron VoiceChat speech-decoder modules for the response-generation side of the duplex stack (#15066).
• Added DuplexSTT training and inference for generating agent text from user speech plus text context (#15092).
• Added Duplex/EAR-TTS speech-generation pieces used by VoiceChat, including MagpieTTS decoder integration and semantic-codec training support (#15277, #15524).
• Added the joint NemotronVoiceChat STT + TTS class for validation, offline speech-to-speech inference, and export workflows (#15456).
• Added formatter/reproducibility work for the VoiceChat speech decoder, with faster training and half-precision inference support (#15583).
• Expanded Voice Agent examples for Nemotron Nano v2/v3, MagpieTTS, tool calling, audio logging, and README guidance (#14325, #14704, #15243, #15269, #15318, #15547, #15642).
• Hardened Voice Agent behavior around default Parakeet EOU STT, text aggregation, end-of-bot handling, RTVI messages, empty tools, and logging (#14951, #15068, #15069, #15634).

Note: NemotronVoiceChat class is inference/eval only. Train DuplexSTTModel and DuplexEARTTS/speech-decoder modules separately.

ASR

• Unified prompt-model support now covers multilingual ASR and streaming inference (#15666).
• Streaming ASR gained Unified RNNT inference, batched streaming beam search, simulated chunked transducer decoding, streaming speech translation, and Canary streaming policies (#15522, #15753, #15517, #15132, #14765).
• Transducer decoding is faster and more complete: 2.4x speedup with confidence, buffered/cache-aware confidence, fewer redundant decoding passes, lower streaming memory, and WER fixes (#15315, #15765, #15301, #15148).
• Phrase boosting expanded across transducers, cache-aware RNN-T, CTC/RNN-T/TDT GPU phrase boosting, and cache-aware customization (#15125, #15344, #14277, #14757, #14800).
• New and updated model families include prompt Parakeet Hybrid RNNT/CTC, ASR EOU models, Streaming Sortformer, Canary2 with NFA, and ASR Transformer/Conformer-Transformer encoders (#14561, #14740, #14627, #14121, #15661, #15703).
• Diarization/VAD removed the Pyannote dependency and gained more flexible input handling (#15632, #15184).

TTS

• MagpieTTS received a major refresh: decoder model, refactor, longform inference, unified longform/standard paths, and updated model internals (#15277, #15504, #15210, #15241, #15477, #15375, #15031).
• Added CFG distillation, local transformer CFG distillation, and MagpieTTS MoE support (#15568, #15748, #15370).
• Expanded language/text support: Hindi, Japanese, Arabic char tokenizer, Japanese-English katakana, hi-IN/ko-KR/pt-BR IPA, and Japanese G2P accent support (#15320, #15248, #15614, #15567, #15170).
• Improved EasyMagpie zero-shot disabling and speaker-encoder behavior (#15639, #15503, #15564).
• Codec and evaluation work added semantic-codec training, codec conversion/bandwidth extension, HF audio-codec loading, Frechet Codec Distance, parallel scoring, comparison reports, and deterministic multi-GPU eval (#15524, #15191, #15172, #15223, #15417, #15621, #15427).
• Inference paths now support reference-free inference, dataset selection, clearer inference config, and fp16 Whisper loading (#15213, #15212, #15254, #15680).

Audio

• Added and fixed Conformer U-Net speech-enhancement models (#14442, #14626).
• Added a data-prediction objective for flow-matching speech enhancement (#14749).
• Added streaming mode to SpectrogramToAudio (#14524).
• Fixed BNR 2.0 inference alignment with padded input signals (#15388).
• Reduced duplication and applied small fixes in the audio collection (#15587).

DataLoader and Speech Data Tools

• Lhotse support expanded to Parquet/Arrow embedded audio, AIS batch loading, AIS-hosted input configs, non-tarred S3 audio, temperature re-weighting, randomized tarred shard slicing, and in-manifest channel selection (#15303, #15102, #15538, #14891, #15732, #15200, #14558, #14586).
• Audio-codec data loading gained Lhotse training format support plus follow-up fixes (#15622, #15742).
• Online data augmentation now covers clipping, lowpass, lossy codec augmentation, and saving augmented audio from Lhotse samplers/dataloaders (#14809, #14808).
• Speech Data Explorer gained S3 reading, comparison mode, security fixes, and tutorial updates (#15500, #15137).

Docs

• Documentation was rewritten around speech scope, uv installs, ASR/TTS/SpeechLM2 guides, and cleaner navigation.

Repo Rename and 3.0 Release

• Project identity is now NVIDIA NeMo Speech in NVIDIA-NeMo/Speech (#15783, #15788).
• README/docs now describe the speech-focused repo, current status, and nightly docs (#15127, #15217, #15602, #15766).

Technical Debt Reduction

• Removed LLM, VLM, diffusion, NLP, multimodal, ExportDeploy, evaluator, Automodel, and language-modeling paths (#14095, #14192, #14617, #14677, #14790, #14934, #14964, #15033, #15044, #15357, #15378).
• Cleaned unsupported ASR/TTS modules, old tutorials, old buffered CTC scripts, and deprecated SpeechLM1 scripts (#14660, #15061, #15368, #15507, #15655).
• Removed unsafe pickle paths and hardened target/export handling (#14540, #15065, #15232, #15266, #15288, #15629).

Installation, Packaging, and Containers

• Switched to a uv-first source install: uv sync --extra all --extra cu13 (#15769).
• Added clear bring-your-own Python/PyTorch/CUDA guidance via uv pip or pip, addressing common install feedback (#15769).
• Routed optional SpeechLM2 compiled deps through uv-compatible sources and container builds: Transformer Engine, FlashAttention, Mamba, grouped GEMM, DeepEP (#15737, #15758).
• Refreshed Docker around official PyTorch bases, lighter staged builds, CUDA 12/13 support, and CVE dependency updates (#15638, #15747, #15756, #15761).
• Updated CI/publish flows to install from uv.lock and simplify release builds (#15659, #15668, #15685, #15697).

Tests and Release Hardening

• Added functional init/train-step/inference tests for every supported released model (#15433).
• Added golden-value MoE dispatch tests for grouped GEMM/expert routing (#15698).
• Split L0 tests into ASR GPU/CPU and SpeechLM2 buckets (#15654).
• Hardened streaming ASR, Sortformer, MagpieTTS, codec, and TTS L2/e2e tests (#14416, #14417, #14435, #14823, #15272, #15508, #15584, #15607).
• Refactored release CI and moved to AWS ephemeral runners (#15620, #15668, #15718).

Detailed PR Breakdown

Release Identity and Docs

• #15788 releases r3.0.0; #15783 bumps the next release to 3.0.
• #15127 renames docs/title to NVIDIA NeMo Speech; #15217 and #15602 update repo status/release messaging.
• #15363 and #15460 restructure docs; #15542 refactors ASR docs; #15647 refactors diarization docs; #15745 improves docs visuals.
• #15769 rewrites install docs for uv, bring-your-own stacks, CUDA extras, and pip fallback.
• #15773 documents SpeechLM para...

NVIDIA Neural Modules 2.7.3

Highlights

• This release addresses known security issues. For the latest NVIDIA Vulnerability Disclosure Information visit https://www.nvidia.com/en-us/security/, for acknowledgement please reach out to the NVIDIA PSIRT team at PSIRT@nvidia.com

Uncategorized:

Changelog

• beep boop 🤖: Bumping nemo to v2.7.3 by @nemo-automation-bot[bot] :: PR: #15552
• remove experimental sclite file by @nithinraok :: PR: #15572
• fix: Remove nemo_class arg in import_hf.py by @chtruong814 :: PR: #15596
• fix: Use RestrictedUnpickler in nemo.lightning.io.mixin by @chtruong814 :: PR: #15597
• cp: fix: Remove diskcache from requirements (15630) into r2.7.0 by @chtruong814 :: PR: #15635

NVIDIA Neural Modules 2.7.2

ASR

Changelog

• cp: Fix numba-cuda and cuda-python installation and usage (#15506) by @chtruong814 :: PR: #15540

Uncategorized:

Changelog

• beep boop 🤖: Bumping nemo to v2.7.2 by @nemo-automation-bot[bot] :: PR: #15535

NVIDIA Neural Modules 2.7.1

ASR

Changelog

• cp: Fix cuda-python usage for CUDA graphs (#15416) by @ko3n1g :: PR: #15471

Uncategorized:

Changelog

NVIDIA Neural Modules 2.7.0

Highlights

• Speech
  • Adds Per-Stream Phrase Boosting in ASR Decoding (Transducers) #15125
  • Adds support for streaming speech translation #15132
  • Released new model nemotron-speech-streaming-en-0.6b that performs English Streaming ASR
  • Released new TTS model magpie_tts_multilingual_357m for multilingual Text-to-Speech

Starting with the next release, NeMo 2.8.0, the following collections will be removed: avlm, diffusion, llm, multimodal, multimodal-autoregressive, nlp, speechlm, vision, vlm, and this repo will focus solely on speech tasks: ASR, TTS, speaker diarization, and speech enhancement.

Detailed Changelogs:

ASR

Changelog

• Enable CUDA graphs in streaming tests by @artbataev :: PR: #14953
• Update ctc-segmentation by @chtruong814 :: PR: #14991
• check asr models by @nithinraok :: PR: #14989
• Unified inference of streaming ASR by @naymaraq :: PR: #14817
• Update numba to numba-cuda and update cuda python bindings usage by @chtruong814 :: PR: #15018
• Fixing lines for multispeaker pipeline by @tango4j :: PR: #15030
• Inference optimization for cache-aware pipelines by @naymaraq :: PR: #15035
• fix loading of hyb ctc rnnt bpe models when using from pretrained by @nithinraok :: PR: #15042
• removed old buffered CTC script by @naymaraq :: PR: #15061
• remove nlp related notebooks by @nithinraok :: PR: #15070
• Update MagpieTTS model with latest changes by @blisc :: PR: #15031
• ASR inference: expose RNN-T decoding params for context biasing by @artbataev :: PR: #15091
• update notebook by @nithinraok :: PR: #15093
• Fix: Obsolete Attribute [SDE] by @Jorjeous :: PR: #15105
• Upgrade NeMo ASR tutorials from Mozilla/CommonVoice to Google/FLEURS by @KunalDhawan :: PR: #15103
• Add support for AIS batch loading for ASR audio processing by @gaikwadabhishek :: PR: #15102
• Multi-Talker Parakeet Streaming - NeMo Documents and Tutorial Notebooks PR 03 by @tango4j :: PR: #15025
• [Fix] Fix the notebook errors on multispeaker data simulation and end to end diarization training by @tango4j :: PR: #15149
• Streaming transducer inference: fix memory usage, improve WER by @artbataev :: PR: #15148
• Execute with subprocess list by @nithinraok :: PR: #15165
• Chunking fix by @nune-tadevosyan :: PR: #15163
• ASR Decoding: allow fallback to CUDA graphs without while loops by @artbataev :: PR: #15173
• remove nlp/modules by @dimapihtar :: PR: #14934
• Asr numpy 2 fix by @nithinraok :: PR: #15166
• Adding flexible input sources for Diarization Mixin by @tango4j :: PR: #15184
• Add support for streaming speech translation by @naymaraq :: PR: #15132
• Confidence fix get_correct_marks by @nune-tadevosyan :: PR: #15128
• Chunking edge cases by @nune-tadevosyan :: PR: #15182
• update subprocess cmd by @nithinraok :: PR: #15218
• Changes required for enabling prompt based models in Nemo Inference by @arushidNV :: PR: #15036
• Fixing the missing sample_rate argument in mixin calling in Sortformer model file by @tango4j :: PR: #15228
• Fix audio tensor loading canary2 by @nithinraok :: PR: #15265
• Fix word confidence return by @nithinraok :: PR: #15249
• feat(asr): add optional auxiliary timestamp model restoration for Canary by @chaosido :: PR: #15268
• Performance: Optimize .nemo tar extraction & model config processing by @paulirish :: PR: #15245
• fix speech commands notebook by @nithinraok :: PR: #15290
• fix timestamps processing with audio tensor input by @nithinraok :: PR: #15291
• Update conv_asr.py preventing unnecessary calculations by @tamilselvan0x0 :: PR: #15239
• Bump to pytorch 25.11 by @chtruong814 :: PR: #15247
• Add FeatureBuffer support to Cache-Aware streaming pipeline by @arushidNV :: PR: #15188
• Per-Stream Phrase Boosting in ASR Decoding (Transducers) by @artbataev :: PR: #15125
• Sort audio by duration in ASR streaming inference script by @artbataev :: PR: #15297
• ASR transcribe: fix forced decoder reinstantiation with timestamps=True by @artbataev :: PR: #15298
• Removes use of torchaudio and moves transforms inside of NeMo by @blisc :: PR: #15211
• Add sacrebleu to ASR requirements by @pzelasko :: PR: #15016
• SpeechLM2 : Add support for offset key in Multimodal conversation by @AudranBert :: PR: #15281
• Add cross-attention to output hypotheses by @mgaido91 :: PR: #15229
• Add warm-ups for RTFX calculation in streaming ASR pipelines by @naymaraq :: PR: #15313
• Speedup buffered transducer inference: remove double decoding by @artbataev :: PR: #15301
• improve canary performance on short audio by @nithinraok :: PR: #15317
• Transducer Decoding: Move fusion models to the base class by @artbataev :: PR: #15322
• Add typing to speech_to_text_finetune.py by @Garvys :: PR: #15326
• Bugfix: correct fusion scores for TDT by @artbataev :: PR: #15325
• Fix ASR streaming script: correctly add biasing requests to model by @artbataev :: PR: #15334
• Fix ASR context biasing in streaming TDT decoding by @artbataev :: PR: #15327

TTS

Changelog

• Remove HeteronymClassificationModel by @blisc :: PR: #14980
• remove nlp.parts collection by @dimapihtar :: PR: #14617
• Update MagpieTTS model with latest changes by @blisc :: PR: #15031
• remove nlp/modules by @dimapihtar :: PR: #14934
• [TTS] MagpieTTS Inference Refactoring by @subhankar-ghosh :: PR: #15178
• [DRAFT][TTS] Magpietts Simple API and loading audiocodec from Huggingface by @subhankar-ghosh :: PR: #15172
• [TTS][MagpieTTS] Change French tokenizer to use 'french_chartokenizer' by @subhankar-ghosh :: PR: #15205
• Add Japanese g2p katakana accent support by @quapham :: PR: #15170
• [TTS][MagpieTTS] Longform TTS using MagpieTTS by @subhankar-ghosh :: PR: #15210
• [voice agent] Fixing the missing arguments calling in NemoSTTService by @SangwonSUH :: PR: #15233
• [TTS] MagpieTTS inference: Add command line option to select a subset of datasets to run inference on by @rfejgin :: PR: #15212
• [TTS] Allow inference without reference audio by @rfejgin :: PR: #15213
• [TTS] Refactor Magpie to support codec conversion and bandwidth extension by @rlangman :: PR: #15191
• [TTS] MagpieTTS: Implement Frechet Codec Distance metric + some minor inference bugfixes by @rfejgin :: PR: #15223
• Update MagpieTTS' Inference Parameter Configuration by @blisc :: PR: #15254
• [TTS][MagpieTTS] Add longform capability to do_tts method by @subhankar-ghosh :: PR: #15241
• [TTS] Add tests of the MagpieTTS inference CLI by @rfejgin :: PR: #15272
• [MagpieTTS][TTS] Support local transformer in longform magpietts by @subhankar-ghosh :: PR: #15296
• Removes use of torchaudio and moves transforms inside of NeMo by @blisc :: PR: #15211
• [MagpieTTS][Docs] Add magpietts docs by @subhankar-ghosh :: PR: #15302
• Add Hindi (hi-IN) support for TTS by @quapham :: PR: #15248
• build: Explicitly set torch >= 2.6.0 and remove weights_only=False by @chtruong814 :: PR: #15314
• [MagpieTTS] Fix incorrect sort order comment in pareto_rank function by @matteolippi :: PR: #15333

NLP / NMT

Changelog

• remove nlp.parts collection by @dimapihtar :: PR: #14617
• chore: remove ExportDeploy by @pablo-garay :: PR: #15033
• remove nlp related notebooks by @nithinraok :: PR: #15070
• Add deprecation notice to modules by @chtruong814 :: PR: #15050
• [OMNIML-3034] ModelOpt rename from TRT ModelOpt to ModelOpt by @yueshen2016 :: PR: #15147
• remove nlp/modules by @dimapihtar :: PR: #14934
• Add support for streaming speech translation by @naymaraq :: PR: #15132
• Remove hardcoded DEBUG logging level in gpt_oss.py by @yurekami :: PR: #15236
• Docs: replace removed preprocess_data_for_megatron.py with Megatron-L… by @Saibabu7770 :: PR: #15222
• remove nlp documentation by @dimapihtar :: PR: #15304
• fix speech translation vllm dockerfile by @naymaraq :: PR: #15310

Text Normalization / Inverse Text Normalization

Changelog

• Add import guards for mcore lightning module by @chtruong814 :: PR: #14970
• chore: update Lightning requirements version by @liquor233 :: PR: #15004

NeMo Tools

Changelog

• Fix: Obsolete Attribute [SDE] by @Jorjeous :: PR: #15105
• Updated tutorial on SDE, due to changes in colab and libraries by @Jorjeous :: PR: #15137

Export

Changelog

• chore: remove ExportDeploy by @pablo-garay :: PR: #15033
• [OMNIML-3034] ModelOpt rename from TRT ModelOpt to ModelOpt by @yueshen2016 :: PR: #15147
• fix: Raise exception in nemo.export instead of allowing pickle.loads by @chtruong814 :: PR: #15266

Bugfixes

Changelog

• Fix PEFT resume with resume_from_path by @maanug-nv :: PR: #14966
• Update deprecated env var by @maanug-nv :: PR: #14975
• Revert lhotse patch after updating to lhotse 1.32.2 by @chtruong814 :: PR: #15329

Uncategorized:

Changelog

• Version bump to 2.7.0rc0.dev0 by @github-actions[bot] :: PR: #14956
• Update changelog for v2.5.1 by @github-actions[bot] :: PR: #14967
• Bump MCore, TE, Pytorch, and modelopt for 25.11 by @chtruong814 :: PR: #14946
• Remove code related to nemo-evaluator (aka nemo-eval) by @athitten :: PR: #14964
• Update changelog for r2.5.0 by @github-actions[bot] :: PR: #14990
• Add clear resharding message error message by @mikolajblaz :: PR: #14962
• Fix Evo2 checkpoint backward compatibility by @farhadrgh :: PR: #14914
• Pass timeout when running speech functional tests by @chtruong814 :: PR: #15012
• [Voice Agent] Fix text aggregation, eob handlin...

NVIDIA Neural Modules 2.6.2

Highlights

• This release addresses known security issues. For the latest NVIDIA Vulnerability Disclosure Information visit https://www.nvidia.com/en-us/security/, for acknowledgement please reach out to the NVIDIA PSIRT team at PSIRT@nvidia.com
• Update tutorial on SDE and imports in Audio Notebook

Detailed Changelogs:

Uncategorized:

Changelog

• cp: Updated tutorial on SDE, due to changes in colab and libraries (15137) into r2.6.0 by @chtruong814 :: PR: #15289
• chore: Bump version to 2.6.2 by @chtruong814 :: PR: #15335
• cp: unset weights_only=False (15312) into r2.6.0 by @chtruong814 :: PR: #15328
• cp: Explicitly set torch >= 2.6.0 and remove weights_only=False (#15314) to r2.6.0 by @chtruong814 :: PR: #15337
• cp: Update Imports in Audio Notebook (15345) into r2.6.0 by @chtruong814 :: PR: #15346
• cp: Clarify when to use TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD (15353) into r2.6.0 by @chtruong814 :: PR: #15358

NVIDIA Neural Modules 2.6.1

Highlights

• This release addresses known security issues. For the latest NVIDIA Vulnerability Disclosure Information visit https://www.nvidia.com/en-us/security/, for acknowledgement please reach out to the NVIDIA PSIRT team at PSIRT@nvidia.com
• Enable numpy 2.0 for ASR

Detailed Changelogs:

ASR

Changelog

• cp: Remove nlp module by @chtruong814 :: PR: #15258
• cp: Enable numpy 2.0 for ASR by @chtruong814 :: PR: #15259

TTS

Changelog

• cp: Remove nlp module by @chtruong814 :: PR: #15258

NLP / NMT

Changelog

• cp: Remove nlp module by @chtruong814 :: PR: #15258

Export

Changelog

• cp: fix: Raise exception in nemo.export instead of allowing pickle.loads (#15266) by @chtruong814 :: PR: #15275

Uncategorized:

Changelog

• Bump version to 2.6.1 by @chtruong814 :: PR: #15244
• cp: guard trust_remote_code (#15065) by @chtruong814 :: PR: #15256
• cp: Execute with subprocess list (#15165) by @chtruong814 :: PR: #15257
• cp: replace pcikle.loads with json.loads (#15232) by @chtruong814 :: PR: #15262
• cp: update subprocess cmd (#15218) by @chtruong814 :: PR: #15263
• cp: [Fix] Fix safety issue for fsdp_dtensor (#15227) by @chtruong814 :: PR: #15276

NVIDIA Neural Modules 2.6.0

Highlights

• Speech
  • Add Timestamps to streaming ASR PR
  • Add Streaming decoding policies (Wait-K and AlignAtt) for Canary model PR
  • Add NeMo Voice Agent PR
  • Hybrid RNNT-CTC Prompted Parakeet Model support PR
  • [New] MT-Parakeet Streaming Models release
• Removed the Automodel module. Automodel is available in the repo https://github.com/NVIDIA-NeMo/Automodel.
• Removed the Deploy module. Export & Deploy is available in the repo https://github.com/NVIDIA-NeMo/Export-Deploy.
• Non-Speech NeMo 2.0 collections are deprecated and will be removed in a later release. Their functionality is available in the Megatron Bridge repo at https://github.com/NVIDIA-NeMo/Megatron-Bridge.

Known Issues

• NeMo voice agent pipecat connecting issues

Detailed Changelogs:

ASR

Changelog

• fixing kernel restarting when transcribing by @weiqingw4ng :: PR: #14665
• Downgrade "datasets" library version in ASR tutorial to ensure compatibility with HF Datasets used by @KunalDhawan :: PR: #14679
• Fixing Sortformer training tutorial notebook by @tango4j :: PR: #14680
• Fix for "EncDecRNNTBPEModel transcribe() failed with TypeError" by @andrusenkoau :: PR: #14698
• Force activations and weights cast to FP32 Jasper Encoder Squeeze-Excite (merge to main) by @erastorgueva-nv :: PR: #14743
• Use lhotse dataloader for ASR models to support in-manifest channel selection for multichannel recordings by @racoiaws :: PR: #14586
• add transducer timestamps without alignments, timestamps to streaming by @lilithgrigoryan :: PR: #14766
• Adding bf16 Sortformer train and inference by @tango4j :: PR: #14627
• Replace texterrors with kaldialign library by @andrusenkoau :: PR: #14775
• fix: Use shutil.copy fallback to handle file metadata permission errors by @vipnydav :: PR: #14639
• Add Customization Capabilities to Cache-Aware Models by @artbataev :: PR: #14757
• Documentation for gpu-based phrase boosting by @andrusenkoau :: PR: #14800
• Streaming decoding policies (Wait-K and AlignAtt) for Canary model by @andrusenkoau :: PR: #14765
• Add tests for streaming buffered and cache-aware transducer models by @artbataev :: PR: #14823
• Merge updates of Multi-Talker Parakeet Model, Modules, Dataloader and Utils PR 01 by @weiqingw4ng :: PR: #14905
• Merge updates of Multi-Talker Parakeet - Unit tests and CI tests PR 02 by @weiqingw4ng :: PR: #14932
• Add Parakeet Hybrid RNNT CTC BPE Model with Prompt support by @ealbasiri :: PR: #14561
• fix notebooks by @nithinraok :: PR: #15079
• cherry pick #15070 by @nithinraok :: PR: #15082

TTS

Changelog

• Remove outdated TTS Tutorials by @blisc :: PR: #14660
• Add KokoroTTS support for voice agent framework by @tango4j :: PR: #14910
• remove language_modeling by @dimapihtar :: PR: #14192

NLP / NMT

Changelog

• Add gpt-oss by @cuichenx :: PR: #14457
• Fix sequence packing loss calculation by @rayandasoriya :: PR: #14437
• [Perf script] Llama and GPT3 perf script use mlp cast fusion by @guyueh1 :: PR: #14575
• Delete tutorials/llm/llama/biomedical-qa directory by @cuichenx :: PR: #14653
• Add gpt-oss lora exporter by @cuichenx :: PR: #14589
• Replace MegatronTokenizer with MegatronLegacyTokenizer by @chtruong814 :: PR: #14721
• Update ModelCommPGs API from megatron-core by @yaoyu-33 :: PR: #14578
• feat: Compatibility modification of megatron-fsdp by @shjwudp :: PR: #14593
• imported get_moe_layer_wise_logging_tracker from megatron core moe_utils by @prathamk-tw :: PR: #14694
• Fix gpt-oss yarn_original_max_position_embeddings value by @cuichenx :: PR: #14706
• Update docs per guidance by @pablo-garay :: PR: #14841
• Fixing three mcore links by @aschilling-nv :: PR: #14839
• Documentation for gpu-based phrase boosting by @andrusenkoau :: PR: #14800
• Update gpt-oss configs by @cuichenx :: PR: #14674
• remove language_modeling by @dimapihtar :: PR: #14192
• cp: remove ExportDeploy into r2.6.0 by @pablo-garay :: PR: #15053
• cherry pick #15070 by @nithinraok :: PR: #15082

Export

Changelog

• fix: fix missing rope scaling in exporting llama embedding model by @ZhiyuLi-Nvidia :: PR: #14523
• Add gpt-oss lora exporter by @cuichenx :: PR: #14589
• Skip trt-llm and vllm install in install test by @chtruong814 :: PR: #14663
• Fix deepseek export dtype by @cuichenx :: PR: #14307
• Remove export-deploy, automodel, and eval tutorials by @chtruong814 :: PR: #14790
• cp: remove ExportDeploy into r2.6.0 by @pablo-garay :: PR: #15053

Uncategorized:

Changelog

• Version bump to 2.6.0rc0.dev0 by @github-actions[bot] :: PR: #14512
• [Audio]: added conformer U-Net model for SE by @nasretdinovr :: PR: #14442
• hyena/evo2: Make sure to convert to real after fp32 conversion by @antonvnv :: PR: #14515
• Force-set restore path for student in KD mode by @AAnoosheh :: PR: #14532
• Skip PTQ if PTQ model path exists by @jenchen13 :: PR: #14536
• Support QwenVL for inference API by @meatybobby :: PR: #14534
• Hyena: Allow to use unfused RMSNorm + TELinear to restore accuracy and some speed by @antonvnv :: PR: #14542
• [Audio]: added streaming mode to SpectrogramToAudio by @nasretdinovr :: PR: #14524
• Update evo2 defaults so converted checkpoints have the right parameters by @jstjohn :: PR: #14514
• deprecate t0 scripts by @dimapihtar :: PR: #14585
• cfg typo correction by @malay-nagda :: PR: #14588
• [Perf script] Add use_te_activation_func and activation_func_fp8_input_store flags by @guyueh1 :: PR: #14522
• Modify logging message to signal that RestoreConfig will be used by @balvisio :: PR: #14469
• Bump TE and Mcore by @chtruong814 :: PR: #14568
• Avoid host-device sync in PTL logging by @WanZzzzzz :: PR: #14489
• Integrate implicit filter kernel with Hyena layer by @farhadrgh :: PR: #14621
• Fix kv_channels configuration for Gemma2 27b by @ananthsub :: PR: #14590
• [Flux] small fixes by @CarlosGomes98 :: PR: #14333
• [Flux] Add MXFP8 Support by @alpha0422 :: PR: #14473
• Use hugginface_hub for downloading the FLUX checkpoint by @suiyoubi :: PR: #14638
• Fine-tune embedding models (E5-Large-V2 and LLaMA-3.2-1B) on the allnli triplet dataset with NeMo Framework by @girihemant19 :: PR: #14584
• remove service launch scripts by @dimapihtar :: PR: #14647
• Warn instead of error when chat template doesn't contain generation keyword by @jenchen13 :: PR: #14641
• Fix function calling notebook by @cuichenx :: PR: #14643
• [Audio]: fixed bug in conformer unet by @nasretdinovr :: PR: #14626
• Fix code checkout during test by @chtruong814 :: PR: #14658
• Fix Flux seed as optional Arg by @suiyoubi :: PR: #14652
• Remove PEFT scheme condition from recipe by @JRD971000 :: PR: #14661
• Add NeMo Voice Agent by @stevehuang52 :: PR: #14325
• Update get_tensor_shapes function whose signature was refactored by @AAnoosheh :: PR: #14594
• Delete nemo1 notebooks by @cuichenx :: PR: #14677
• Bump latest Mcore 020abf01 by @chtruong814 :: PR: #14676
• [Flux] correct vae_downscale_factor by @CarlosGomes98 :: PR: #14425
• Bump modelopt to 0.35.0 and remove safe_import("modelopt") in llm collection by @kevalmorabia97 :: PR: #14656
• Canary tutorial fix by @nune-tadevosyan :: PR: #14699
• Add option for LoRA with Transformer Engine op fuser by @timmoon10 :: PR: #14411
• add load-in-4bit param by @dimapihtar :: PR: #14636
• Support NVFP4 recipe by @WanZzzzzz :: PR: #14625
• Fix broken link in Reasoning-SFT.ipynb by @cuichenx :: PR: #14716
• Remove artificial block to vortex fp8 TP by @jstjohn :: PR: #14684
• Drop speech_llm example suite by @yaoyu-33 :: PR: #14683
• remove env var by @malay-nagda :: PR: #14739
• detach arg option for run scripts by @malay-nagda :: PR: #14722
• Randomized shard slicing for tarred data by @pzelasko :: PR: #14558
• Data prediction objective for flow matching speech enhancement models by @racoiaws :: PR: #14749
• Fix Some Failures by @alpha0422 :: PR: #14763
• Support additional Slurm parameters (#14701) by @bdubauski :: PR: #14742
• [Flux] Remove Redundant Host & Device Sync by @alpha0422 :: PR: #14711
• [Flux] Full Iteration CUDA Graph by @alpha0422 :: PR: #14744
• Update prune-distill notebooks to Qwen3 + simplify + mmlu eval by @kevalmorabia97 :: PR: #14785
• ci: Automodel deprecation warning by @thomasdhc :: PR: #14787
• Bug in MXFP8 recipe by @adityavavreNVDA :: PR: #14793
• feat: Disable blank Issues by @pablo-garay :: PR: #14788
• ci: Add community label bot by @chtruong814 :: PR: #14796
• Add mistral small3 24B config and recipe by @eagle705 :: PR: #14784
• Update changelog for r2.3.0 by @github-actions[bot] :: PR: #14812
• QWEN2.5-VL 7B FP8 Recipe by @tomlifu :: PR: #14801
• Feat: Disk space management: for nemo install test by @pablo-garay :: PR: #14822
• Evo2 address rare over-masking in 1m context dataset by @jstjohn :: PR: #14821
• Update cherry-pick workflow to use version 0.63.0 by @pablo-garay :: PR: #14832
• Removing automodel items by @aschilling-nv :: PR: #14840
• Update changelog for v2.4.1 by @github-actions[bot] :: PR: #14828
• Fix lm_eval installation in pruning tutorial for 25.09 container by @kevalmorabia97 :: PR: #14865
• Add nemotron-nano-v2 support to voice agent by @stevehuang52 :: PR: #14704
• Update changelog for 2.5.0 by @chtruong814 :: PR: #14890
• [Qwen3] Fix the flop cal for Qwen3 by @gdengk :: PR: #14897
• [lhotse][aistore] added support input_cfg.yaml directly from aistore bucket by @XuesongYang :: PR: #14891
• Harden _is_target_allowed by adding runtime class validation on top of prefix checks to prevent unsafe target resolution by @KunalDhawan :: PR: #14540
• Enable simplifi...

NVIDIA Neural Modules 2.5.3

Highlights

• This release addresses known security issues. For the latest NVIDIA Vulnerability Disclosure Information visit https://www.nvidia.com/en-us/security/, for acknowledgement please reach out to the NVIDIA PSIRT team at PSIRT@nvidia.com
• Update nv-one-logger
• Update ctc-segmentation

Detailed Changelogs:

Text Normalization / Inverse Text Normalization

Changelog

• chore: update Lightning requirement by @liquor233 :: PR: #15005

Uncategorized:

Changelog

• cp: Update ctc-segmentation (14991) into r2.5.0 by @chtruong814 :: PR: #15020
• Bump to 2.5.3 by @chtruong814 :: PR: #15022

NVIDIA Neural Modules 2.5.2

Detailed Changelogs:

Text Normalization / Inverse Text Normalization

Changelog

• cp: Add import guards for mcore lightning module (#14970) into r2.5.0 by @chtruong814 :: PR: #14982

Uncategorized:

Changelog

• Bump to 2.5.2 by @chtruong814 :: PR: #14983