How to transform the timestamps output by the model to seconds?

[My-captain]

How to transform the timestamps output by the model to seconds?

[My-captain]

看了源码，请问是否÷1000就可以了？

[LauraGPT]

Yes. For the standard FunASR AutoModel.generate() result, timestamps are in milliseconds, so divide both boundaries by 1000.0.

For character/word timestamps:

result = model.generate(input="audio.wav")
item = result[0]

timestamps_seconds = [
    [start_ms / 1000.0, end_ms / 1000.0]
    for start_ms, end_ms in item.get("timestamp", [])
]

For VAD/speaker sentence segments:

sentences_seconds = [
    {
        **segment,
        "start": segment["start"] / 1000.0,
        "end": segment["end"] / 1000.0,
    }
    for segment in item.get("sentence_info", [])
]

For example, [1230, 2480] becomes [1.23, 2.48] seconds. Keep the original millisecond values if you need integer subtitle or alignment boundaries, and only convert fields whose contract is milliseconds.

Current source references:

• AutoModel result contract: [[start_ms, end_ms], ...]
• README sentence example using start / 1000
• Timestamp postprocessing stores integer milliseconds