Error with Canary Model: "Computing timestamps are not supported for this model yet"
Hello,
Thank you for sharing the Canary model (nvidia/canary-1b-flash). I am currently using it for audio transcription and have encountered an issue when attempting to generate timestamps. I receive the following error:
Computing timestamps are not supported for this model yet.
Here is the code I am using:
from nemo.collections.asr.models import EncDecMultiTaskModel
canary_model = EncDecMultiTaskModel.from_pretrained('nvidia/canary-1b-flash')
decode_cfg = canary_model.cfg.decoding
decode_cfg.beam.beam_size = 1
canary_model.change_decoding_strategy(decode_cfg)
output = canary_model.transcribe(
['filepath.wav'],
timestamps='yes', # This line causes the error
)
predicted_text = output[0].text
word_level_timestamps = output[0].timestamp['word']
segment_level_timestamps = output[0].timestamp['segment']
The model should generate timestamps at both the word and segment levels.
Actual Behavior:
The model throws an error indicating that timestamp computation is not supported.
Questions:
Is timestamp generation supported in the nvidia/canary-1b-flash model? If not, are there any plans to add this feature in the future?
Are there alternative models or configurations within the NVIDIA NeMo framework that support timestamp generation for audio transcription?
Thank you for your assistance. Please let me know if you need additional details or logs to investigate this issue.
Best regards,
Karim
Hi Karim,
Could you check if you are using the latest NeMo main branch to run the above inference code? Timestamp support for Canary-Flash models was added in this PR. Please let me know if this doesn’t resolve the issue.
Hi Kunal,
Thanks for checking! I confirm everything is working well on my end with the latest NeMo main branch—timestamp support works perfectly with the Canary-Flash models after pulling the recent updates.
Merci 🙏