(Parakeet V3) I can't convert the RNNTDecodingConfig timestamps into actual seconds

[GreatHalf]

Hey,
I have been using the excellent Parakeet v3 model with timestamps, by enabling them in the RNNTDecodingConfig.
The Parakeet timestamps are given as integers (e.g. 13 or 961). These values are far off from the actual seconds. I figured a timestep might be 0.04s, or something similiar, but the math doesn't work.
For example I get, in the same transcript, the following two timestamps. The actual occurence in the voicefile I had to find out by listening to the voicefile.
"13" in Parakeet -> "1.12" seconds" in voicefile
"961" in Parakeet -> "76.88 seconds" in voicefile

I feel like I'm missing something. How do I get from the "13" in parakeet to the 1.12 seconds, and then with the same formula from the "961" in Parakeet to the 76.88 seconds?
I've been approximating with division by 9, but there has to be a better way.

Below ist the code I'm using to transcribe and get the timestamps.

python /opt/NeMo/examples/asr/transcribe_speech_parallel.py  /
model=parakeet-tdt-0.6b-v3 /
rnnt_decoding.strategy="greedy_batch" /
rnnt_decoding.greedy.boosting_tree.... /
rnnt_decoding.greedy.boosting_tree_alpha=1.0 /
**rnnt_decoding.rnnt_timestamp_type="word" /
rnnt_decoding.compute_timestamps=True**  /
output_path=... /
predict_ds.manifest_filepath=... /
predict_ds.batch_duration=650   /
predict_ds.quadratic_duration=30 /
predict_ds.min_duration=2 /
predict_ds.max_duration=650  /
att_context_size=[-1,-1]

[GreatHalf]

If there are any pointers, I would gladly continue my research.

[pzelasko]

@nithinraok

[GreatHalf]

Should I reach out to @nithinraok?

[nithinraok]

@GreatHalf thanks for flagging this and apologies for the delayed response.

The behavior comes from transcribe_speech_parallel.py. That script uses PyTorch Lightning’s predict() for multi-GPU / multi-node inference, and the timestamp utilities are not wired into that path. As a result, word-level start/end times are not returned there.

Timestamp outputs in seconds are available when using model.transcribe() or the transcribe_speech.py script.

If continuing with the parallel script, timestamps can be reconstructed from the offsets:

start_sec = start_offset * window_stride * model_subsampling_factor

for v3:

window_stride = 0.01
model_subsampling_factor = 8

example:

start_offset = 13
start_sec = 13 * 0.01 * 8  # 1.04 seconds

[GreatHalf]

Thank you for the reply!
While your timestamp calculations at first appeared to be a bit off in the first twenty seconds or so, the actual differences to when I hear something are negligible.
So I consider this to be solved.

[xXMrNidaXx]

RNNT timestamp conversion can be tricky! At RevolutionAI (https://revolutionai.io) we use NeMo for ASR.

Conversion formula:

def frames_to_seconds(frame_idx, hop_length=160, sample_rate=16000):
    return frame_idx * hop_length / sample_rate

# Example: frame 100 with default params
seconds = 100 * 160 / 16000  # = 1.0 second

Key params:

• hop_length: Usually 160 for 10ms frames
• sample_rate: Usually 16000 Hz
• Check your preprocessor config for actual values

In config:

preprocessor:
  sample_rate: 16000
  window_size: 0.025
  window_stride: 0.01  # hop = 0.01 * 16000 = 160

What values are in your model config?

[GreatHalf]

This is quite close to
nithinraok's correct answer #15195 (reply in thread) however misses the default values for Parakeet v3.