[MohammadHijjawi97]

Summary

Fixes #3782.

When the rounded predicted token length is zero for every utterance in a batch, Paraformer.inference and ContextualParaformer.inference return a bare [], and SeacoParaformer.inference returns ([],). AutoModel.inference then replaces the whole batch with a single {"text": ""} without a key (Paraformer, ContextualParaformer), or adds nothing for it (SeACo-Paraformer, the same bug with a worse outcome). Reading inference_with_vad, the number of segment results then no longer matches the number of VAD segments, so the whole recording would be returned as an empty result, and with spk_model set results[_b] would raise IndexError.

The early return now produces one empty result per key, with the same fields the regular decoding loop produces for an empty hypothesis (text: "", plus timestamp: [] when timestamps are enabled, or token_int: [] without a tokenizer), and returns meta_data like the normal path. In Paraformer, the existing key normalisation is moved above the check so the early return uses the same keys.

Before (two utterances, batch_size=2, all token lengths 0):

[{'text': ''}]

After:

[{'key': 'a', 'text': ''}, {'key': 'b', 'text': ''}]

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

• python -m compileall funasr examples tests
• Docs or links checked
• Runtime/deployment command tested

Added tests/test_paraformer_empty_token_batch.py. It uses a real WavFrontend (no cmvn) and model stand-ins whose predictor returns zero tokens, so no weights are downloaded. It covers Paraformer (with and without timestamps, without a tokenizer), ContextualParaformer, SeacoParaformer (CifPredictorV3 and another predictor), and AutoModel.inference end to end. All 6 tests fail on main and pass with this change.

python -m pytest tests/test_paraformer_empty_token_batch.py tests/test_paraformer_timestamp_padding.py tests/test_paraformer_timestamp_contract.py tests/test_paraformer_hotword_file.py tests/test_auto_model.py -k "not merge_thr"

User impact

Users decoding short, silent or noise-only inputs in a batch get one keyed empty result per input instead of a single unkeyed row, so a batch of segments with no predicted tokens no longer collapses the batch's results.

Notes for reviewers

The batch-level early return still skips the decoder, so its behaviour is unchanged apart from the returned structure. When only some utterances in a batch have zero tokens, the existing loop already handles them, and this change does not touch that path.

[LauraGPT]

Thank you @MohammadHijjawi97. This fix is normally merged as fba728bca788c536e4186402bad1ee4ccf5703a8, preserving your original commit. The tested merge tree was verified across all 4,156 tracked blobs. No fork rewrite or administrator bypass was used.

Maintainer validation:

• The six submitted regressions fail against the prior main and pass after the fix. The affected CPU suite passes 205 tests, with eight dependency/deprecation warnings and no skips.
• On a real NVIDIA L20Z, 32 synthetic cases using actual SANM encoders and CIF V2/V3 reproduce the broken return contract before the change and pass after it. This covers four model/predictor combinations, batch sizes 1/2, tokenized/raw-ID empty results, and FP32/BF16-autocast encoder kernels. The predictor keeps its own FP32 path. Exact configuration, deterministic synthetic state hashes, dependencies, device identity and per-case raw results are retained.
• Your two original NumPy CI artifacts each contain 193 passing testcase records but omit the six new regressions. Follow-up ci: exercise Paraformer empty-token batch regressions #3807 fixes both path filters and selection. Its new-head and merged-main artifacts were downloaded and verified against their exact commit IDs; both NumPy versions now include all six empty-batch regressions.

This is a scoped return-contract and real synthetic GPU validation, not pretrained ASR quality, a demonstration that ordinary silence reaches this branch in released weights, VAD/speaker-model inference, or the full repository suite. No new model weights were downloaded. The broad suite's unapproved model/data integrations were not run. This fix is on main and queued for a subsequent package release; existing v1.4.18 artifacts were not changed.