Add 🤗 Transformers usage to model card

README.md

@@ -47,6 +47,7 @@ datasets:
 - voxpopuli
 - europarl
 tags:
+- transformers
 - speech-recognition
 - cache-aware ASR
 - automatic-speech-recognition
@@ -366,6 +367,8 @@ pip install git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]
 
 The model is available for use in the NeMo Framework, and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
 
+You can also run it with [🤗 Transformers](https://github.com/huggingface/transformers) (more below).
+
 ### Loading the Model
 
 ```python
@@ -407,6 +410,125 @@ Latency is defined by the `att_context_size` param, where att_context_size = `{n
 
 Here, chunk size = current frame + right context; each chunk is processed in non-overlapping fashion.
 
+### 🤗 Transformers usage
+
+This checkpoint also runs with [🤗 Transformers](https://github.com/huggingface/transformers). The target language is passed through the processor's `language` argument: 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; it is a special token, so decoding with `skip_special_tokens=True` strips it (clean transcript) and `skip_special_tokens=False` keeps it for language labeling.
+
+Until Nemotron3_5Asr is part of an official Transformers release, install Transformers from source:
+
+```bash
+pip install git+https://github.com/huggingface/transformers
+```
+
+<details>
+  <summary>➡️ Offline transcription</summary>
+
+```python
+from transformers import AutoModelForRNNT, AutoProcessor
+from transformers.audio_utils import load_audio
+
+model_id = "nvidia/nemotron-3.5-asr-streaming-0.6b"
+revision = "refs/pr/20"
+processor = AutoProcessor.from_pretrained(model_id, revision=revision)
+model = AutoModelForRNNT.from_pretrained(model_id, revision=revision, device_map="auto")
+
+audio = load_audio(
+    "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
+    sampling_rate=processor.feature_extractor.sampling_rate,
+)
+
+# Condition on a known language ...
+inputs = processor(audio, sampling_rate=processor.feature_extractor.sampling_rate, language="en-US")
+inputs.to(model.device, dtype=model.dtype)
+output = model.generate(**inputs, return_dict_in_generate=True)
+print(processor.decode(output.sequences, skip_special_tokens=True))
+
+# ... or let the model detect it and keep the emitted <xx-XX> language tag.
+inputs = processor(audio, sampling_rate=processor.feature_extractor.sampling_rate, language="auto")
+inputs.to(model.device, dtype=model.dtype)
+output = model.generate(**inputs, return_dict_in_generate=True)
+print(processor.decode(output.sequences, skip_special_tokens=False))
+```
+</details>
+
+<details>
+  <summary>➡️ Streaming transcription</summary>
+
+```python
+from threading import Thread
+from transformers import AutoModelForRNNT, AutoProcessor, TextIteratorStreamer
+from transformers.audio_utils import load_audio
+
+model_id = "nvidia/nemotron-3.5-asr-streaming-0.6b"
+revision = "refs/pr/20"
+processor = AutoProcessor.from_pretrained(model_id, revision=revision)
+model = AutoModelForRNNT.from_pretrained(model_id, revision=revision, device_map="auto")
+
+processor.set_num_lookahead_tokens(6)
+print(f"Streaming latency: {processor.streaming_latency_ms} ms")
+
+# The language prompt rides along on every chunk; use a locale (e.g. "de-DE") or "auto".
+language = "en-US"
+
+sampling_rate = processor.feature_extractor.sampling_rate
+audio = load_audio(
+    "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/obama.mp3",
+    sampling_rate=sampling_rate,
+)
+
+first_chunk_inputs = processor(
+    audio[: processor.num_samples_first_audio_chunk],
+    sampling_rate=sampling_rate,
+    is_streaming=True,
+    is_first_audio_chunk=True,
+    language=language,
+    return_tensors="pt",
+)
+first_chunk_inputs = first_chunk_inputs.to(model.device, dtype=model.dtype)
+
+
+def input_features_generator():
+    yield first_chunk_inputs.input_features[:, : processor.num_mel_frames_first_audio_chunk, :]
+
+    mel_frame_idx = processor.num_mel_frames_first_audio_chunk
+    hop_length = processor.feature_extractor.hop_length
+    n_fft = processor.feature_extractor.n_fft
+
+    start_idx = mel_frame_idx * hop_length - n_fft // 2
+    while (end_idx := start_idx + processor.num_samples_per_audio_chunk) < audio.shape[0]:
+        inputs = processor(
+            audio[start_idx:end_idx],
+            sampling_rate=sampling_rate,
+            is_streaming=True,
+            is_first_audio_chunk=False,
+            language=language,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(model.device, dtype=model.dtype)
+        yield inputs.input_features
+
+        mel_frame_idx += processor.num_mel_frames_per_audio_chunk
+        start_idx = mel_frame_idx * hop_length - n_fft // 2
+
+
+streamer = TextIteratorStreamer(processor.tokenizer, skip_special_tokens=True)
+generate_kwargs = {
+    **first_chunk_inputs,
+    "input_features": input_features_generator(),
+    "streamer": streamer,
+}
+thread = Thread(target=model.generate, kwargs=generate_kwargs)
+thread.start()
+
+print("Model output (streaming):", end=" ", flush=True)
+for text_chunk in streamer:
+    print(text_chunk, end="", flush=True)
+thread.join()
+```
+</details>
+
+For more details about usage, please refer to the [Transformers documentation](https://huggingface.co/docs/transformers/en/model_doc/nemotron3_5_asr).
+
 ### Input(s): <br>
 
 **Input Type(s):** Audio, Lang ID <br>