nvidianemotron-3.5-asr-streaming-0.6b
Nemotron¶
Uses Nemotron's cache-aware native decoder and requests word timestamps.
Parameter metadata: Exact learned-parameter total for VoiceHub's audited native primary graph at the registered default selection; separately loaded auxiliary models are excluded.
Usage¶
Complete the VoiceHub installation once, then run this repository-authored example. Model pages intentionally contain no package-install command.
This example is maintained against VoiceHub's public API; it is not copied from an upstream demo or package README.
Model-specific path: Uses Nemotron's cache-aware native decoder and requests word timestamps.
Inputs and controls: Chunk geometry is owned by the checkpoint runtime; common chunk and stride overrides intentionally fail closed.
from pathlib import Path
from voicehub import AutoModelForSpeechRecognition
AUDIO_FILE = Path("speech.wav")
if not AUDIO_FILE.is_file():
raise FileNotFoundError(AUDIO_FILE)
model = AutoModelForSpeechRecognition.from_pretrained(
'nvidia/nemotron-3.5-asr-streaming-0.6b',
model_type='asr_nemotron',
device="cuda",
lazy_load=True,
)
output = model.transcribe(
AUDIO_FILE,
return_timestamps="word",
)
print(output.text)
for segment in output.segments:
print(segment.start, segment.end, segment.text, segment.confidence)
Use authorized recordings. Verify hardware needs and pin a revision in production.
Overview¶
asr_nemotron is a VoiceHub automatic speech recognition
integration. This page is generated from its registry contract. Open the asr_nemotron Colab notebook.
| Property | Value |
|---|---|
| Task | Automatic speech recognition |
| Architecture | nemotron-3.5-rnnt |
| Runtime | VoiceHub-native |
| Languages | en-US, en-GB, es-US, es-ES, … complete audited list below |
| Capabilities | automatic-speech-recognition, multilingual, language-identification, timestamps, streaming-architecture, safetensors, fine-tuning, voicehub-native, native-runtime |
| Reusable components | — |
| Normalized output | ASROutput |
Language support¶
Supported language abbreviations
en-US, en-GB, es-US, es-ES, fr-FR, fr-CA, it-IT, pt-BR, pt-PT, nl-NL, de-DE, tr-TR, ru-RU, ar-AR, hi-IN, ja-JP, ko-KR, vi-VN, uk-UA, pl-PL, sv-SE, cs-CZ, nb-NO, da-DK, bg-BG, fi-FI, hr-HR, sk-SK, zh-CN, hu-HU, ro-RO, et-EE, el-GR, lt-LT, lv-LV, mt-MT, sl-SI, he-IL, th-TH, nn-NO
The el-GR, lt-LT, lv-LV, mt-MT, sl-SI, he-IL, th-TH, and nn-NO locales are adaptation-ready and require in-domain fine-tuning; the other listed locales are transcription-ready or broad-coverage.
Paper and GitHub¶
- Paper: No dedicated upstream research paper is published for this integration.
- Upstream GitHub: NVIDIA NeMo
- VoiceHub source: VoiceHub model implementation
Configuration¶
Load configuration without constructing the model:
from voicehub import AutoConfig
config = AutoConfig.for_model('asr_nemotron')
print(config.model_type)
| Property | Value |
|---|---|
| Canonical model type | asr_nemotron |
| Configuration class | NemotronASRConfig |
| Architecture class | NemotronForSpeechRecognition |
Processing¶
Create the registered processor without allocating model weights:
from voicehub import AutoProcessor
processor = AutoProcessor.from_pretrained(
'nvidia/nemotron-3.5-asr-streaming-0.6b',
model_type='asr_nemotron',
)
print(type(processor).__name__)
Inference¶
The Usage example returns ASROutput through AutoModelForSpeechRecognition.
Input and output contract¶
| Property | Value |
|---|---|
| Readiness | integrated-raw |
| Data architecture | rnnt |
| Sample rate | 16,000 Hz |
| Contract getter | get_asr_dataset_spec('asr_nemotron') |
| Variant | Required fields | One of | Boundary | Other rules |
|---|---|---|---|---|
raw-audio |
audio |
text / transcription / transcript | Source | at most one: text / transcription / transcript |
nemotron-rnnt-model-ready |
input_features, attention_mask, prompt_ids, labels, label_lengths, decoder_input_ids |
— | Prepared | — |
Language-prompted Nemotron RNN-T fine-tuning records. See the data workflow.
Training and optimization¶
Use available_optimization_passes() to discover reversible public passes.
Unsupported runtime or hardware fails closed before mutation.
Training contract¶
| Property | Value |
|---|---|
| Support | native |
| Family | rnnt |
| Recipe | single-phase |
| Default phase | speech_recognition |
| Training checkpoint | nvidia/nemotron-3.5-asr-streaming-0.6b |
| Native training graph | yes |
| Phase | Kind | Components | Required inputs | Loss keys |
|---|---|---|---|---|
speech_recognition |
objective | model.encoder, model.encoder_projector, model.prompt_projector, model.decoder, model.joint |
input_features, attention_mask, prompt_ids, labels, label_lengths, decoder_input_ids |
loss |
The integration accepts its declared source or prepared contract directly. Call model.validate_training_support() first, then follow the
training workflow.
Checkpoints, provenance, license, and limitations¶
| Property | Value |
|---|---|
| Default checkpoint | nvidia/nemotron-3.5-asr-streaming-0.6b |
| Hugging Face ID | nvidia/nemotron-3.5-asr-streaming-0.6bRepository availability verified through the Hugging Face model API on 2026-08-11; pin a revision before production use. |
| Checkpoint status | Registry default; pin an immutable revision for production and reproducible evidence |
| Optional dependency extra | Core package |
| Hardware and runtime | Usage selects cuda; verify checkpoint-specific requirements |
| Real-checkpoint evidence | Release evidence; a registry default alone is not execution evidence |
| Implementation | voicehub.models.asr_nemotron.modeling_asr_nemotron.NemotronForSpeechRecognition |
| Configuration | voicehub.models.asr_nemotron.configuration_asr_nemotron.NemotronASRConfig |
| Source provenance | voicehub/architectures/nemotron_asr/SOURCE.json |
| License | OpenMDW-1.1 |
Use of the checkpoint and derivatives is governed by the OpenMDW-1.1 license. Commercial use: allowed by the registered terms.
Confirm the checkpoint revision, access terms, provenance, and license.
Limitations¶
- No integration-specific checkpoint limitation is registered. Verify the selected checkpoint revision and its documented runtime requirements.
- Validate memory, precision, and optional dependencies on the target system.
- Public optimizations fail closed when the runtime or hardware cannot satisfy their validation contract; an unavailable pass is not reported as applied.
- Contract tests do not replace the linked released-checkpoint evidence.
Public API¶
Use the stable configuration, processor, and task-model facades below.
Configuration
NemotronASRConfig¶
Parameters¶
**config_kwargs— Configuration fields validated by NemotronASRConfig.
Model
NemotronForSpeechRecognition¶
Parameters¶
pretrained_model_name_or_path— Hub ID or compatible local directory.model_type— Canonical model type; use 'asr_nemotron'.config— Optional preloaded NemotronASRConfig instance.**model_kwargs— Model-specific loading arguments.
from voicehub import get_model_spec
spec = get_model_spec('asr_nemotron')
print(spec.display_name, spec.task.value)
| Purpose | Public object |
|---|---|
| Discover | get_model_spec('asr_nemotron') |
| Load and run | AutoModelForSpeechRecognition |
| Configure | NemotronASRConfig |
| Process | AutoProcessor |
| Model implementation | NemotronForSpeechRecognition |
| Normalized output | ASROutput |
| Training contract | get_training_spec('asr_nemotron') |
| Optimization lifecycle | available_optimization_passes, apply_optimization_plan, optimization_manifest, restore_optimization_plan |
See all model guides, inference, and the training matrix.