Audio: add local duration-aware ASR batching

benchmarking/scripts/audio_librispeech_benchmark.py

@@ -177,6 +177,9 @@ def run_audio_librispeech_benchmark(  # noqa: PLR0913
                 model_id=model_name,
                 audio_filepath_key="audio_filepath",
                 batch_size=asr_batch_size,
+                max_audio_sec_per_actor=240.0,
+                max_inference_duration_s=120.0,
+                local_bucketing=True,
                 fail_on_audio_error=True,
                 adapter_kwargs={"use_cuda_graph_decoder": False},
             )

nemo_curator/models/asr/base.py

@@ -70,6 +70,7 @@ class ASRAdapter(Protocol):
     * ``waveform``: contiguous, mono, 1-D float32 NumPy samples normalized by
       ``ASRStage`` from a file or a reader-provided in-memory waveform.
     * ``sample_rate`` (``int``): the stage's configured target sample rate.
+    * ``audio_seconds`` (``float``): duration of this model-input chunk.
     * ``language`` (``str | None``): human-readable name (e.g. ``"English"``).
     * ``language_code`` (``str | None``): original language code from the
       configured stage input column.