A Real-Time Wideband Neural Vocoder at 1.6 kb/s Using LPCNet

[jmvalin]

This is a follow-up on the first LPCNet demo. In this new demo, we turn LPCNet into a very low-bitrate neural speech codec (see submitted paper) that's actually usable on current hardware and even on phones. It's the first time a neural vocoder is able to run in real-time using just one CPU core on a phone (as opposed to a high-end GPU). The resulting bitrate — just 1.6 kb/s — is about 10 times less than what wideband codecs typically use. The quality is much better than existing very low bitrate vocoders and comparable to that of more traditional codecs using a higher bitrate.

Read More

Comments

Crash on raspi

[(Anonymous)]

Hello,

Thank you for the new codec. I have tested it on my desktop without avx2 and like the quality/space tradeoff. It's painfully slow decoding without avx2, but that is expected. I tried to cross-compile for raspberry pi 3 and get a crash in sincos. Do you know what I'm doing wrong? I'm using gcc-8.3.0 and I had to do a bit of work to get it to compile:
CC=arm-unknown-linux-gnueabi-gcc-8.3.0 CFLAGS='-O3 -ggdb -march=armv8-a -mfpu=neon' ../configure --target=arm-unknown-linux-gnueabi --host=x86_64-pc-linux-gnu
I'm on commit 343e35.

Re: Crash on raspi

[jmvalin]

Looks like a toolchain problem. Why not build directly on the Raspberry Pi instead of cross-compiling. If you really want to cross-compile, start with just a hello-world like with a main that does "printf("%f\n", sin(time(NULL)));"

About the code being slow without AVX2, if you don't have an AVX2-capable CPU, it's probably not very hard to convert the AVX2 code into SSE with very little performance loss. I haven't benchmarked, but I suspect even a Core 2 should be able to run LPCNet in real-time with SSE code.