Bugs in MossFormer2_SE_48K direct NumPy inference for audio longer than 20 seconds

[f0sta1]

Hello! I found several bugs in the direct NumPy/tensor-to-tensor inference path for MossFormer2_SE_48K in clearvoice/utils/decode_batch.py.

I use the official NumPy API:

audio = np.reshape(audio, [1, audio.shape[0]]).astype(np.float32)

model = ClearVoice(
    task="speech_enhancement",
    model_names=["MossFormer2_SE_48K"]
)

output = model(audio, False)

Short audio works correctly, but audio longer than one_time_decode_length (20 seconds) enters the segmented inference branch and exposes several issues.

1. Incorrect input length calculation

The input has shape [batch, samples], but the code uses:

input_len = inputs.shape[0]

For an input of shape (1, 9240399), this sets input_len to 1 instead of 9240399.

Fix:

input_len = inputs.shape[-1]

2. Incorrect indexing of the output tensor

outputs is created with shape [batch, samples]:

outputs = torch.from_numpy(np.zeros((b, t)))

but later it is indexed with three indices:

outputs[batch_idx, batch_idx, ...]

Fix:

outputs[batch_idx, ...]

This applies to both output assignments in the segmented inference loop.

3. Padded output is not trimmed to the original input length

For a 60-second, 48 kHz mono input:

Input shape:  (1, 2880000)
Output shape: (1, 3168000)

The 60-second input therefore produced 66 seconds of output.

Fix:

outputs = outputs[:, :input_len]

before the final return.

Results after the fixes

After these three fixes, a 60-second file completed successfully with the exact original output length:

INPUT duration=60.000 sample_rate=48000 samples=2880000 dtype=float32
NUMPY input_shape=(1, 2880000) dtype=float32
TIMING inference=74.087
NUMPY output_shape=(1, 2880000) dtype=float64
TIMING total=75.633

A longer file that previously caused the Python process to terminate during inference with exit code 9 and empty stderr also completed successfully after the fixes:

INPUT duration=192.508 sample_rate=48000 samples=9240399 dtype=float32
NUMPY input_shape=(1, 9240399) dtype=float32
TIMING inference=321.712
NUMPY output_shape=(1, 9240399) dtype=float64
TIMING total=323.386

Before the fixes, the same 192.508-second file failed during the direct NumPy inference call with:

Exit code: 9
stderr:

Exact diff

@@ -334,7 +334,7 @@
-    input_len = inputs.shape[0]  # Get the length of the input audio
+    input_len = inputs.shape[-1]  # Get the number of audio samples

@@ -408,10 +408,10 @@
-                        outputs[batch_idx, batch_idx, current_idx:current_idx + window - give_up_length] = output_segment[:-give_up_length]
+                        outputs[batch_idx, current_idx:current_idx + window - give_up_length] = output_segment[:-give_up_length]
                     else:
                         output_segment = output_segment[-window:]
-                        outputs[batch_idx, batch_idx, current_idx + give_up_length:current_idx + window - give_up_length] = output_segment[give_up_length:-give_up_length]
+                        outputs[batch_idx, current_idx + give_up_length:current_idx + window - give_up_length] = output_segment[give_up_length:-give_up_length]

@@ -476,6 +476,9 @@
+    # Trim padded output back to the exact original input length.
+    outputs = outputs[:, :input_len]
+
     return outputs.numpy() / MAX_WAV_VALUE

Environment

* macOS on Apple Silicon
* Python 3.11
* Model: MossFormer2_SE_48K
* Input: mono WAV, 48 kHz, float32
* API mode: direct NumPy/tensor-to-tensor inference through ClearVoice

Thank you for making ClearerVoice-Studio available. I hope these findings are useful.