Deepspeech: Interested in benchmarks (client)

[jacobjennings]

Preface: I'm new to ML, go easy :)

I'm very exited to hear that large trained models will be published soon. I'm curious about one thing, since I'm working with constrained hardware. For a larger trained model (600-1000 Mb), I'd love to see a benchmark of "seconds of speech processed per second" on any system as a baseline. Particularly, I'm interested to see if a DeepSpeech client is in the realm of viability on something like an iPhone for offline speech applications.

Thanks to all contributors!

[kdavis-mozilla]

Our goal is to allow for off-line speech applications, e.g. iPhone ot Raspberry Pi 3. As part of that we're working on compressing models using quantization (32 bits to 8 bits) and SVD.

However, it's a bit early for us to start publishing benchmarks as we're still in the process of finalizing our architecture and means of quantization.

[captainnurple]

Hi @kdavis-mozilla, curious what your ROUGH timeline estimate is for off-line recognition esp. on mobile device. Thanks!

[kdavis-mozilla]

@captainnurple Very rough, next 3-6 months.

[ghost]

@kdavis-mozilla any news on this one? thanks

[saurabhvyas]

@kdavis-mozilla Very curious how much latency you get on rpi 3, and do you have a docker image for it, so we can play with it ?

[lock[bot]]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.