[tanishqbaweja]

Summary

Fixes #3881

ContextualParaformer.proc_hotword in funasr_onnx and funasr_torch (SeacoParaformer inherits it) split the hotword string with split(" ") and padded each hotword into a row of 10 ids with pad_list(..., max_len=10):

• A hotword longer than 10 characters did not fit in its row, so the whole call raised ValueError: could not broadcast input array from shape (12,) into shape (10,) (for 12 characters).
• Repeated, leading or trailing spaces, and an empty hotword string, gave empty hotwords. Each became a row of ten 0 ids whose embedding was read at position -1 (the last of the 10), and went to the model as an extra hotword.

The exported embedding model takes exactly 10 ids per hotword: ContextualEmbedderExport (funasr/models/seaco_paraformer/export_meta.py, subclassed for the Contextual Paraformer export) makes only num_hotwords dynamic, so model_eb.onnx declares its input as ['num_hotwords', 10]. The C++ runtime keeps the first max_hotword_len = 10 tokens of each hotword and skips hotwords without tokens (CompileHotwordEmbedding in runtime/onnxruntime/src/paraformer.cpp:565-607 and paraformer-torch.cpp:530-572).

This PR does the same in proc_hotword (three lines of code and a two-line comment per file):

• split() instead of split(" "), as generate_hotwords_list splits in funasr/models/seaco_paraformer/model.py (:688) and funasr/models/contextual_paraformer/model.py (:653). This drops the empty strings.
• Each hotword keeps its first max_hotword_len = 10 characters, and pad_list uses the same constant. These wrappers map one character to one id, so that is 10 ids, as in C++.

What proc_hotword returns (each hotword's ids without padding, <s> = 1, and the positions where __call__ reads the embeddings), with the toy vocabulary of the script in the issue:

hotwords	main	this PR
'魔搭 达摩院'	[[14, 15], [6, 7, 8], [1]], [1, 2, 0]	unchanged
'魔搭 达摩院'	[[14, 15], [], [6, 7, 8], [1]], [1, -1, 2, 0]	[[14, 15], [6, 7, 8], [1]], [1, 2, 0]
''	[[], [1]], [-1, 0]	[[1]], [0]
'魔搭 阿里巴巴达摩院语音实验室' (12 characters)	ValueError	[[14, 15], [3, 4, 5, 5, 6, 7, 8, 9, 10, 11], [1]], [1, 9, 0]

Type of change

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

Validation

I added test_hotword_wrappers_skip_empty_and_cut_long_hotwords to tests/test_runtime_eval_native.py, for ContextualParaformer and SeacoParaformer in funasr_onnx and funasr_torch. It loads each wrapper with its real constructor from the hotword fixture graphs of test_hotword_wrappers_decode_every_batch (whose model_eb.onnx has the same dynamic axes as the export). It checks the hotword ids and embedding positions for " 魔搭 你好魔搭你好魔搭你好魔搭 " and "", and that both strings decode all three fixture files.

test-runtime-eval.yml runs this file for changes under runtime/python/onnxruntime/funasr_onnx/** and runtime/python/libtorch/funasr_torch/**, so no workflow change is needed. With the pins it installs (Python 3.12.15, torch 2.8.0+cpu, numpy 1.26.4, onnx 1.19.0, onnxruntime 1.23.2), on Windows 11:

python -m pytest -q tests/test_runtime_eval_cli.py tests/test_runtime_eval_native.py

• New test on main source (0e52c3c): 4 failed, 55 passed. All four fail with ValueError: could not broadcast input array from shape (12,) into shape (10,).
• This branch: 59 passed.
• This branch merged with the head of fix(runtime): trim batch padding from CIF peaks before funasr_onnx and funasr_torch Paraformer timestamps #3871, which also adds tests to this file: 63 passed.

Also:

• The same command with Python 3.11.17, numpy 2.4.6, torch 2.10.0+cpu, onnxruntime 1.23.2: 4 failed, 55 passed on main source; 59 passed on this branch.

• tests/test_rich_transcription_postprocess.py, tests/test_rich_transcription_event_segments.py and tests/test_paraformer_hotword_file.py: 40 passed.

• The funasr_onnx package scripts pass from source: the release contract (with PYTHONUTF8=1 on Windows), overlap, empty_final, vad_online_short_chunks, vad_multiple_inputs and punctuation_empty_input. test_funasr_onnx_installed.py passes 9 of 10; the other one reads the metadata of an installed wheel.

• tests/test_funasr_torch_package.py: 13 passed, 1 failed, the same as on main. The failure is test_wheel_contains_all_runtime_sources, which on Windows builds its expected file names with backslashes.

• I exported a tiny embedder (random embedding and LSTM) with ContextualEmbedderExport itself, using its dummy input, input/output names and dynamic axes, and opset 14 (the default in funasr/utils/export_utils.py). ONNX Runtime runs a [3, 10] hotword matrix and rejects [3, 12] with Got invalid dimensions for input: hotword. So padding the rows to the longest hotword instead would fail with an exported model.

• black --line-length=100 (24.4.0, as in pre-commit) leaves the test file unchanged and does not touch the changed lines of the two paraformer_bin.py files (it would reformat other lines of those files, on main too).

• python -m compileall funasr examples tests

• Docs or links checked

• Runtime/deployment command tested (the wrappers with the synthetic fixture graphs, CPU)

No real model was run.

User impact

Users of funasr_onnx or funasr_torch ContextualParaformer and SeacoParaformer can pass hotwords longer than 10 characters: the model gets their first 10 characters, as with the C++ runtime, instead of the whole call failing. Extra spaces between, before or after hotwords, and an empty hotword string, no longer add a padding-only hotword. Hotword strings with single spaces and hotwords of up to 10 characters give the same model input as before.

Notes for reviewers

• The cut is silent, as in the C++ runtime. I can add a logging.warning for cut hotwords if you prefer.
• split() also splits on tabs and newlines, as generate_hotwords_list does. split(" ") kept them inside a hotword, where they became <unk> ids with an OOV warning. The C++ runtime splits on spaces only.
• Characters outside the vocabulary are still mapped to <unk> with a warning (the C++ runtime skips such hotwords). This PR does not change that.
• git merge-tree --write-tree of this branch with the heads of fix(runtime): trim batch padding from CIF peaks before funasr_onnx and funasr_torch Paraformer timestamps #3871 (which also edits both paraformer_bin.py files and the test file), Fix/security hardening #3877, fix(runtime): honor use_itn in funasr_onnx and funasr_torch SenseVoiceSmall #3860 and fix(runtime): honor SenseVoice ITN and preserve mixed batch query tags #3861 gives no conflicts.
• No version bump and no workflow change.