Add Nemotron 3.5 ASR Streaming

docs/source/en/model_doc/nemotron3_5_asr.md

@@ -0,0 +1,74 @@
+<!--Copyright 2026 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+*This model was contributed to Hugging Face Transformers on 2026-06-24.*
+
+# Nemotron3_5Asr
+
+## Overview
+
+Nemotron3_5Asr is the **multilingual** extension of [NemotronAsr](./nemotron_asr) (`nvidia/nemotron-3.5-asr-streaming-0.6b`).
+It reuses the entire cache-aware streaming [Fast Conformer](https://huggingface.co/papers/2305.05084) encoder, RNN-T
+(Recurrent Neural Network Transducer) head, feature extraction, and streaming generation of [`NemotronAsr`], and adds
+**language-ID prompt conditioning** so a single model transcribes 40 language-locales.
+
+The target language is turned into a one-hot vector, broadcast across the encoder time axis, concatenated with the
+encoder output, and fused back to the encoder hidden size by a small MLP (`prompt_kernel`) before the joint network.
+Pass the language through the processor's `language` argument (Whisper-style; a locale such as `"en-US"`/`"de-DE"`, a
+bare code such as `"de"`, or `"auto"` for automatic language detection). In `auto` mode the model appends an `<xx-XX>` language tag
+after the transcript's terminal punctuation. The tag is a special token, so `decode`/`batch_decode` with the default
+`skip_special_tokens=True` strip it (clean transcript); pass `skip_special_tokens=False` to keep it for language labeling.
+
+## Usage
+
+```python
+from transformers import AutoProcessor, Nemotron3_5AsrForRNNT
+from datasets import load_dataset, Audio
+
+model_id = "nvidia/nemotron-3.5-asr-streaming-0.6b"
+processor = AutoProcessor.from_pretrained(model_id)
+model = Nemotron3_5AsrForRNNT.from_pretrained(model_id).eval()
+
+ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
+
+# Condition on a known language ...
+inputs = processor(ds[0]["audio"]["array"], sampling_rate=16000, language="en-US")
+generated = model.generate(**inputs)
+print(processor.batch_decode(generated.sequences, skip_special_tokens=True))
+
+# ... or let the model detect it and keep the emitted language tag.
+inputs = processor(ds[0]["audio"]["array"], sampling_rate=16000, language="auto")
+generated = model.generate(**inputs)
+print(processor.batch_decode(generated.sequences, skip_special_tokens=False))
+```
+
+## Nemotron3_5AsrConfig
+
+[[autodoc]] Nemotron3_5AsrConfig
+
+## Nemotron3_5AsrProcessor
+
+[[autodoc]] Nemotron3_5AsrProcessor
+
+## Nemotron3_5AsrRNNTOutput
+
+[[autodoc]] Nemotron3_5AsrRNNTOutput
+
+## Nemotron3_5AsrForRNNT
+
+[[autodoc]] Nemotron3_5AsrForRNNT
+    - forward
+    - generate

src/transformers/models/auto/auto_mappings.py

@@ -402,6 +402,7 @@
         ("mvp", "MvpConfig"),
         ("nanochat", "NanoChatConfig"),
         ("nemotron", "NemotronConfig"),
+        ("nemotron3_5_asr", "Nemotron3_5AsrConfig"),
         ("nemotron_asr_streaming", "NemotronAsrStreamingConfig"),
         ("nemotron_asr_streaming_encoder", "NemotronAsrStreamingEncoderConfig"),
         ("nemotron_h", "NemotronHConfig"),
@@ -1047,6 +1048,7 @@
         ("musicflamingo", "MusicFlamingoProcessor"),
         ("musicgen", "MusicgenProcessor"),
         ("musicgen_melody", "MusicgenMelodyProcessor"),
+        ("nemotron3_5_asr", "Nemotron3_5AsrProcessor"),
         ("nemotron_asr_streaming", "NemotronAsrStreamingProcessor"),
         ("nougat", "NougatProcessor"),
         ("omdet-turbo", "OmDetTurboProcessor"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -49,6 +49,7 @@
         ("moonshine", "Wav2Vec2FeatureExtractor"),
         ("moshi", "EncodecFeatureExtractor"),
         ("musicgen", "EncodecFeatureExtractor"),
+        ("nemotron3_5_asr", "NemotronAsrStreamingFeatureExtractor"),
         ("nemotron_asr_streaming_encoder", "NemotronAsrStreamingFeatureExtractor"),
         ("parakeet_ctc", "ParakeetFeatureExtractor"),
         ("parakeet_encoder", "ParakeetFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -338,6 +338,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):
         ("mvp", "MvpModel"),
         ("nanochat", "NanoChatModel"),
         ("nemotron", "NemotronModel"),
+        ("nemotron3_5_asr", "Nemotron3_5AsrForRNNT"),
         ("nemotron_asr_streaming", "NemotronAsrStreamingForRNNT"),
         ("nemotron_asr_streaming_encoder", "NemotronAsrStreamingEncoder"),
         ("nemotron_h", "NemotronHModel"),
@@ -1726,6 +1727,7 @@ class _BaseModelWithGenerate(PreTrainedModel, GenerationMixin):
 MODEL_FOR_RNNT_MAPPING_NAMES = OrderedDict(
     [
         # Model for RNN Transducer (RNN-T) mapping.
+        ("nemotron3_5_asr", "Nemotron3_5AsrForRNNT"),
         ("nemotron_asr_streaming", "NemotronAsrStreamingForRNNT"),
         ("parakeet_rnnt", "ParakeetForRNNT"),
     ]

src/transformers/models/auto/tokenization_auto.py

@@ -233,6 +233,7 @@
         ("musicgen_melody", "T5Tokenizer" if is_tokenizers_available() else None),
         ("mvp", "MvpTokenizer" if is_tokenizers_available() else None),
         ("myt5", "MyT5Tokenizer"),
+        ("nemotron3_5_asr", "ParakeetTokenizer" if is_tokenizers_available() else None),
         ("nemotron_asr_streaming", "ParakeetTokenizer" if is_tokenizers_available() else None),
         ("nezha", "BertTokenizer" if is_tokenizers_available() else None),
         ("nllb", "NllbTokenizer" if is_tokenizers_available() else None),

src/transformers/models/nemotron3_5_asr/__init__.py

@@ -0,0 +1,28 @@
+# Copyright 2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_nemotron3_5_asr import *
+    from .modeling_nemotron3_5_asr import *
+    from .processing_nemotron3_5_asr import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

src/transformers/models/nemotron3_5_asr/configuration_nemotron3_5_asr.py

@@ -0,0 +1,97 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+from ..nemotron_asr_streaming.configuration_nemotron_asr_streaming import NemotronAsrStreamingEncoderConfig
+
+
+@auto_docstring(checkpoint="nvidia/nemotron-3.5-asr-streaming-0.6b")
+@strict
+class Nemotron3_5AsrConfig(PreTrainedConfig):
+    r"""
+    vocab_size (`int`, *optional*, defaults to 13088):
+        Vocabulary size of the joint network output (including the blank token).
+    decoder_hidden_size (`int`, *optional*, defaults to 640):
+        Hidden size of the LSTM prediction network (NeMo's `pred_hidden`).
+    num_decoder_layers (`int`, *optional*, defaults to 2):
+        Number of LSTM layers in the prediction network.
+    hidden_act (`str`, *optional*, defaults to `"relu"`):
+        Activation in the joint network.
+    max_symbols_per_step (`int`, *optional*, defaults to 10):
+        Maximum number of non-blank symbols emitted per encoder time step during greedy decoding.
+    encoder_config (`Union[dict, NemotronAsrStreamingEncoderConfig]`, *optional*):
+        The config object or dictionary of the encoder. Reuses [`NemotronAsrStreamingEncoderConfig`] directly,
+        since the encoder is identical to [`NemotronAsrStreaming`]'s.
+    blank_token_id (`int`, *optional*, defaults to 13087):
+        Blank token id for RNN-T decoding.
+    joint_hidden_size (`int`, *optional*, defaults to 640):
+        Hidden size of the joint network's encoder/decoder projections (NeMo's `joint_hidden`).
+    durations (`list[int]`, *optional*, defaults to `()`):
+        Pinned to the empty tuple for RNN-T: no token durations are predicted, so the joint head outputs
+        only `vocab_size` logits.
+    num_prompts (`int`, *optional*, defaults to 128):
+        Number of language-prompt slots. The target language is encoded as a one-hot vector of this
+        size, broadcast across the encoder time axis and concatenated with the encoder output before
+        the `prompt_kernel` fusion MLP.
+    prompt_intermediate_size (`int`, *optional*, defaults to 2048):
+        Hidden size of the `prompt_kernel` fusion MLP (`Linear(hidden + num_prompts -> intermediate)
+        -> ReLU -> Linear(intermediate -> hidden)`).
+
+    Example:
+    ```python
+    >>> from transformers import Nemotron3_5AsrForRNNT, Nemotron3_5AsrConfig
+
+    >>> configuration = Nemotron3_5AsrConfig()
+    >>> model = Nemotron3_5AsrForRNNT(configuration)
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "nemotron3_5_asr"
+    # The encoder is identical to NemotronAsrStreaming's, so reuse its config class directly.
+    sub_configs = {"encoder_config": NemotronAsrStreamingEncoderConfig}
+
+    vocab_size: int = 13088
+    decoder_hidden_size: int = 640
+    num_decoder_layers: int = 2
+    hidden_act: str = "relu"
+    max_symbols_per_step: int = 10
+    encoder_config: dict | PreTrainedConfig | None = None
+    pad_token_id: int = 0
+    blank_token_id: int = 13087
+    is_encoder_decoder: bool = True
+    joint_hidden_size: int = 640
+    durations: list[int] | tuple[int, ...] = ()
+    num_prompts: int = 128
+    prompt_intermediate_size: int = 2048
+
+    def __post_init__(self, **kwargs):
+        if self.decoder_hidden_size != self.joint_hidden_size:
+            raise ValueError(
+                "Nemotron3_5AsrConfig currently requires decoder_hidden_size == joint_hidden_size "
+                f"(got {self.decoder_hidden_size} and {self.joint_hidden_size})."
+            )
+        # The decoder starts on the blank token at frame 0 (NeMo's blank_as_pad convention).
+        kwargs.setdefault("decoder_start_token_id", self.blank_token_id)
+        if isinstance(self.encoder_config, dict):
+            self.encoder_config = NemotronAsrStreamingEncoderConfig(**self.encoder_config)
+        elif self.encoder_config is None:
+            self.encoder_config = NemotronAsrStreamingEncoderConfig()
+        self.initializer_range = self.encoder_config.initializer_range
+        super().__post_init__(**kwargs)
+
+
+__all__ = ["Nemotron3_5AsrConfig"]