fix(nano): avoid duplicate fs and preserve audio loading errors

[txfs19260817]

Summary

Type of change

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

Validation

Tested on Windows with Python 3.12, using CPU tensors without model weights.

python -m pytest -q tests/test_fun_asr_nano_train_opts.py -k audio_loading
# 12 passed

python -m pytest -q tests/test_fun_asr_nano_train_opts.py -k "not torch_compile_wraps_cuda"
# 19 passed, 2 deselected

python -m compileall -q funasr examples tests

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

User impact

The initial error occurs because fs is supplied both explicitly and through kwargs:

[ERROR] Loading wav failed! funasr.utils.load_utils.load_audio_text_image_video() got multiple values for keyword argument 'fs'

Traceback (most recent call last):
  File ".../funasr/models/fun_asr_nano/model.py", line 551, in data_load_speech
    data_src = load_audio_text_image_video(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: funasr.utils.load_utils.load_audio_text_image_video() got multiple values for keyword argument 'fs'

The handler logs this exception but continues into feature extraction with data_src unassigned, producing a second error that masks the original cause:

Traceback (most recent call last):
  File ".../transcribe.py", line 225, in transcribe_file
    result = model.generate(**generate_kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../funasr/auto/auto_model.py", line 887, in generate
    results = self.inference_with_vad(
              ^^^^^^^^^^^^^^^^^^^^^^^
  File ".../funasr/auto/auto_model.py", line 1109, in inference_with_vad
    results = self.inference(
              ^^^^^^^^^^^^^^^
  File ".../funasr/auto/auto_model.py", line 957, in inference
    res = model.inference(**batch, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../funasr/models/fun_asr_nano/model.py", line 807, in inference
    return self.inference_llm(
           ^^^^^^^^^^^^^^^^^^^
  File ".../funasr/models/fun_asr_nano/model.py", line 967, in inference_llm
    inputs_embeds, contents, batch, source_ids, meta_data = self.inference_prepare(
                                                          ^^^^^^^^^^^^^^^^^^^^^^^
  File ".../funasr/models/fun_asr_nano/model.py", line 657, in inference_prepare
    output = self.data_load_speech(contents, tokenizer, frontend, meta_data=meta_data, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../funasr/models/fun_asr_nano/model.py", line 560, in data_load_speech
    data_src,
    ^^^^^^^^
UnboundLocalError: cannot access local variable 'data_src' where it is not associated with a value

This restores Nano audio loading through the VAD pipeline when fs is forwarded.

Notes for reviewers

The duplicate-keyword pattern dates to the initial Nano integration in 3e2c4c4 (December 2025). Explicit VAD-to-ASR fs forwarding in 9d61129 (October 2026) exposed it in the default VAD path

[LauraGPT]

Validated exact head 2117b036496e2d3dba2d1e8f8a0cd4d50eb2bfc5 against current main a4ab3b9d204f9c8b14cd6d6846abaa29ffff4f7a and synthetic merge be212d45ef3376e56efddd93b39dff3d8fc97988.

• Recursive tree reconciliation found exactly the two Nano model copies and the submitted test file; changed blobs match the head and every other blob matches current main.
• The drifted synthetic merge passes the submitted audio-loading cases (12/12), adjacent Nano option tests (19/19), and changed-file compileall.
• On a real NVIDIA H100 with Torch 2.8.0+cu128/CUDA 12.8, the existing official zh.mp3 through the existing Fun-ASR-Nano-2512 + FSMN-VAD weights reproduces the masked UnboundLocalError(data_src) on current main. The exact synthetic merge returns one non-empty result. CUDA hooks observed 2,691 ASR and 12 VAD Linear calls; both model components and tensors were on cuda:0.

This is a focused one-recording, batch-size-1 regression and hardware validation, not a full model/language/accuracy suite. Within that scope, the fix preserves the forwarded sampling-rate contract and surfaces loader errors instead of continuing with an unbound value.