Suggestion neural-network architecture with small parameter size

[davidlin409]

In current icefall architecture, number of parameters for different architectures are usually large than 1 million parameters. For example:

• TDNN-LSTM: 1-million parameters
• Conformer: 10-million parameters

If somehow in some very small device, which only allows parameter size of ~ 80K parameter, the choice of neural-network architecture seems to be very restricted.

Currently what I can think of is only to use TDNN-F from old Kaldi. Is there any more advanced method which also support small parameter size?

On the other hand, for very small device, is BPE model possible?

• If there are too many output in neural-network, parameter size will also be too much.

[Cirediallo]

Hello,
the subject is closed but can you point out where I can find the receipes architecture details(like how many LSTM layers was used, ...) ?

[csukuangfj]

where I can find the receipes architecture details(like how many LSTM layers was used, ...)

Please have a look at the RESULTS.md in each recipe. You can find the training commands for each pre-trained model.