Can it detect laugh in audio provide timestamps??

[Shivansh-yadav13]

I'm looking for a solution which can detect laugh from audio and provide the timestamps for it.
I came across this project and I'm not sure if it can do that...

[LauraGPT]

Yes for clip-level detection, but not as a native event-level timestamp.

SenseVoiceSmall can emit the raw event tag <|Laughter|>. Check result["text"] before calling rich_transcription_postprocess, because the postprocessor turns that tag into an emoji. For example:

from funasr import AutoModel

model = AutoModel(
    model="iic/SenseVoiceSmall",
    trust_remote_code=True,
    device="cuda:0",  # use "cpu" if needed
)

result = model.generate(
    input="audio.wav",
    language="auto",
    use_itn=False,
    output_timestamp=True,
)[0]

print(result["text"])
print("laughter detected:", "<|Laughter|>" in result["text"])
print(result.get("words"), result.get("timestamp"))

The important limitation is that output_timestamp=True aligns transcription tokens only. The language, emotion, event, and ITN tags are the first four special outputs and are excluded before forced alignment, so <|Laughter|> itself will not appear in words/timestamp.

For approximate laughter locations in a long recording, run SenseVoice on overlapping windows of the raw audio and keep the windows that contain <|Laughter|>:

import numpy as np
import soundfile as sf


def laughter_windows(path, window_seconds=2.0, hop_seconds=0.5):
    audio, sample_rate = sf.read(path, dtype="float32")
    if audio.ndim == 2:
        audio = audio.mean(axis=1)

    original_samples = len(audio)
    window = round(window_seconds * sample_rate)
    hop = round(hop_seconds * sample_rate)
    if original_samples < window:
        audio = np.pad(audio, (0, window - original_samples))

    last_start = len(audio) - window
    starts = list(range(0, last_start + 1, hop))
    if starts[-1] != last_start:
        starts.append(last_start)

    duration = original_samples / sample_rate
    hits = []
    for start in starts:
        raw_text = model.generate(
            input=audio[start : start + window],
            fs=sample_rate,
            language="auto",
            use_itn=False,
        )[0]["text"]
        if "<|Laughter|>" in raw_text:
            hits.append(
                {
                    "start": start / sample_rate,
                    "end": min((start + window) / sample_rate, duration),
                }
            )
    return hits


print(laughter_windows("audio.wav"))

These are window-level, approximate intervals rather than exact event boundaries. In production, merge overlapping hits and usually require two consecutive positive windows to reduce false positives. Tune the window/hop on audio from your own domain; detection quality varies with recording conditions.