Incorrect annotation time intervals after running analyze in terminal

[dyang0910]

Describe the bug
Annotation time intervals aren't 3s after training a customized model and running analyze in the terminal. Instead, we get a time interval that is 3 * speed (the audiospeed we specify in our train parameters).

To Reproduce
Training customized model: python -m birdnet_analyzer.train
“data/train"
--output “/Output“
--test_data "data/test"
--hidden_units 256
--dropout 0.75
--batch_size 64
--learning_rate 0.01
--crop_mode "center"
--overlap 0.5
--threads 8
--fmin 500
--fmax 8000
--audio_speed 10
--upsampling_mode "repeat"
--upsampling_ratio 0.0
--focal-loss-alpha 0.25
--focal-loss-gamma 2.0

Deploy: python -m birdnet_analyzer.analyze "data/test"
--classifier "/CustomClassifier.tflite"
--output “/Output"
--fmin 500
--fmax 8000
--audio_speed 10
--overlap 0.5

Expected behavior
Expected the annotated audio intervals to be 3s, but we get 10s after speeding up the audio 10x. (See screenshot)

Screenshots

Desktop (please complete the following information):
MacBook Air M1, MacOS 15.7.1; Python 3.11; BirdNET V2.4

[max-mauermann]

This is the expected behavior.
BirdNET always processes 3 seconds of audio.

So with a speed-modifier of 10 it would analyse 30s of the original audio as one chunk.

For example:

• a 10 minute recording would be sped up to be 1 minute
• BirdNET then processes this 1 minute in 3s chunks
• the timestamps of the detections are then adjusted to match their position in the original file