funasr_onnx streaming Paraformer repeats the last characters when chunk_size[0] is 0 #3789

Description

funasr_onnx.paraformer_online_bin.Paraformer.add_overlap_chunk keeps the left context for the next chunk with

cache["feats"] = overlap_feats[:, -self.chunk_size[0] :, :]

When chunk_size[0] == 0, -0: is 0:, so the slice returns every frame instead of none. When the final call is split into two chunks, the second chunk is prefixed with all of the first chunk's features, so the tail is decoded twice and the last characters are repeated.

chunk_size=[0, 10, 5] is the streaming setting used in README.md, README_zh.md and docs/python_api*.md, so users following the docs with the ONNX runtime hit this. The package's own demo uses [5, 10, 5] and is not affected. The C++ runtime does not use this code path.

The non-final branch has the same pattern (-(chunk_size[0] + chunk_size[2]):), which only matters when both values are 0.

Reproduction

Environment: funasr and funasr_onnx from main (fbbfe01), onnxruntime 1.20.1, numpy 1.26.4, Python 3.12, CPU.

1. Download and export the model (writes model.onnx and decoder.onnx next to the checkpoint):

from modelscope import snapshot_download
from funasr import AutoModel

model_dir = snapshot_download(
    "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online", revision="v2.0.5"
)
AutoModel(model=model_dir, device="cpu", disable_update=True).export(type="onnx", quantize=True)

2. Optional, to get deterministic output: add dither: 0.0 under frontend_conf: in model_dir/config.yaml. The config does not set it, so the frontend uses its default dither=1.0. The numbers below use 0.0; the duplication also appears with the default.

3. Decode in 9600-sample (600 ms) pieces with chunk_size=[0, 10, 5]:

import soundfile as sf
from funasr_onnx.paraformer_online_bin import Paraformer

model = Paraformer(model_dir, batch_size=1, chunk_size=[0, 10, 5], quantize=False)
speech, _ = sf.read("BAC009S0764W0129.wav", dtype="float32")  # AISHELL-1 test
cache, out, step = {}, [], 9600
n = (len(speech) - 1) // step + 1
for i in range(n):
    res = model(speech[i * step : (i + 1) * step], param_dict={"cache": cache, "is_final": i == n - 1})
    out += [r["preds"][0] for r in res if r]
print("".join(out))

AISHELL-1 test examples:

The single repeat left in W0126 also appears in the PyTorch reference with the same settings, so the fix does not claim to remove every trailing repeat; it only removes the extra decoding of the previous chunk.

On the full AISHELL-1 test set (7176 utterances), 2170 outputs (30.2%) change after the fix, and CER (punctuation removed) goes from 17.23% to 15.18% with the FP32 export and from 17.15% to 15.08% with the quantized export. The PyTorch reference with the same settings is 14.81%.

Proposed fix

Compute the start index explicitly, feats[:, max(feats.shape[1] - n, 0):, :], for both cache updates. PR to follow.