Parakeet v2, v3 max_duration difference

[mingewang]

HI

I noticed a difference in model_config.yaml between Parakeet v2 and Parakeet v3 models:

Parakeet v2
train_ds:
max_duration: 40
validation_ds:
max_duration: 40

Parakeet v3
train_ds:
max_duration: 10
validation_ds:
max_duration: 30

Does this mean that Parakeet v3 was trained with training samples truncated to 10 seconds, while Parakeet v2 was trained with up to 40-second samples?
If so, what was the motivation for reducing the training max_duration in v3?

When I try to increase train_ds.max_duration for Parakeet v3 to 30 seconds, I easily run into CUDA OOM errors on an A100 (80GB).

Are there recommended settings (batch size, gradient accumulation, bucketing strategy, etc.) to safely train v3 with longer utterances?

Any guidance on best practices for handling longer audio with Parakeet v3 ( finetune) would be appreciated.

Thanks!

[jeremy110]

@mingewang
Depending on your dataset, but typically max_duration is set to 30 to 40, and then used with lhotse for dataloader. The relevant settings can be referenced if you're interested. https://github.com/jeremy110/Finetune_Nemo_ASR

[xXMrNidaXx]

Parakeet v2 vs v3 max_duration differences:

Key changes:

Version	Default max_duration	Why
v2	20s	Memory optimization
v3	30s+	Better architecture

Why v3 handles longer:

• Improved attention mechanism
• Better memory efficiency
• Optimized chunking

Adjusting for your needs:

model:
  preprocessor:
    max_duration: 30  # Increase for longer audio
  
  # If OOM, try:
  encoder:
    subsampling_factor: 8  # More aggressive

Memory considerations:

# Estimate memory per second
memory_per_sec = model_params * 4 / 1e9  # GB
max_safe = available_gpu_memory / memory_per_sec

For very long audio:

• Chunk into segments
• Use streaming inference
• Enable gradient checkpointing for training

Config for long-form:

trainer:
  precision: bf16  # Save memory
  accumulate_grad_batches: 4

We train ASR models at RevolutionAI. What's your target audio duration?

[mingewang]

The target audio duration is: 30-40 seconds. are you able to fine-tune/train parakeet v3 ? If so, what is your GPU's spec. e.g: vRAM ? Thanks.