All-zero Paraformer token lengths collapse a batch to one unkeyed result #3782

Source-level report; runtime reproduction is still pending.

🐛 Bug

In the ordinary non-VAD, single-best Paraformer inference path, a batch whose rounded predicted token lengths are all zero returns a bare empty list. AutoModel then substitutes one empty-text result without a key for the whole batch.

I checked current main 879a219c717123fdec6024e8e56c734a7aace2ff. The same early return is present in ContextualParaformer.

To Reproduce

This is a conditional source trace, not an executed audio/model reproduction. No inference command has been run for this report.

1. Consider two inputs with keys ["a", "b"], processed together with batch_size=2, without VAD and with single-best decoding.
2. If their predicted token lengths round to [0, 0], Paraformer lines 618–620 return []. ContextualParaformer lines 458–460 do the same.
3. AutoModel lines 958–963 treat that empty return as [{"text": ""}] and extend the public result list with this single unkeyed row.

Code sample

Relevant current source:

pre_token_length = pre_token_length.round().long()
if torch.max(pre_token_length) < 1:
    return []

The wrapper then uses:

res = model.inference(**batch, **kwargs)
if isinstance(res, (list, tuple)):
    results = res[0] if len(res) > 0 else [{"text": ""}]
    meta_data = res[1] if len(res) > 1 else {}
asr_result_list.extend(results)

Thus, conditional on the all-zero branch, the two-input batch contributes [{"text": ""}], rather than two keyed rows. Processing all-zero items in separate batches would instead contribute one unkeyed row per batch.

Expected behavior

Preserve the two input keys and order, for example [{"key": "a", "text": ""}, {"key": "b", "text": ""}].

The model registration contract specifies a (results, meta_data) model return and retaining ASR input order/keys. The existing batch key test also expects two results with distinct keys for two inputs at batch sizes 1 and 2; that is a routing control, not a reproduction of this predictor branch.

Could you confirm the intended empty-result behavior and model return shape before I work on a focused runtime reproduction and fix?

Error logs

None: source inspection only. I have not demonstrated a real audio input that reaches this branch, and do not claim that silence necessarily triggers it.

Environment

• Source: main commit 879a219c717123fdec6024e8e56c734a7aace2ff
• Python / PyTorch / torchaudio / ModelScope / device: not exercised for this report.
• No model weights, audio, GPU, CLI, or end-to-end inference validation was performed.

Audio details

No audio reproduction yet. This report is limited to the ordinary non-VAD model-return and batching path; other integrations have not been tested.