asumagic

Conformer-transducer streaming model in SpeechBrain

The #2140 PR implements a streaming model in SpeechBrain, based on a Conformer-Transducer architecture.
The "transducer" part refers to (part of) the loss used (RNN-T loss), but the vast majority of changes needed for streaming were specific to the Conformer model.

Two main parameters exist:

asumagic

#include <cstdlib>
#include <cstdio>

int StringEqualUntil(const char* s1, const char* s2, int n)
{
    int c1, c2;

    while (1)
    {
        c1 =  *s1++;

asumagic

#include <algorithm>
#include <iterator>
#include <vector>

// Generator Iterator test
namespace impl
{
template<class Func>
class generator_iterator : public std::iterator<
        std::random_access_iterator_tag,

asumagic

// silly hack job to parse http://www.pastraiser.com/cpu/gameboy/gameboy_opcodes.html into a list with hex and binary representations of opcodes, in order to find patterns.

#include <algorithm>
#include <string>
#include <iostream>
#include <iomanip>
#include <fstream>

int main(int argc, char** argv)
{

asumagic

-- Baseline pineout
-- This is some base pinout file for the C5G, aka the terasic cyclone v gx starter kit board, because the one from terasic is verilog.
-- Some features are just not usable with this file. Don't assume it works or it is reliable.

library ieee;
use ieee.std_logic_1164.all;

entity baseline_c5gx is
	-- 1.2V ADC
	port (

asumagic

[10:40:44] Benchmark: Running 'naive translation (string concat)' (50000000 samples)
[10:43:07] Benchmark: Done in 143 seconds (~0.00286ms per sample)
[10:43:07] Benchmark: Running 'replace() translation' (50000000 samples)
[10:46:11] Benchmark: Done in 184 seconds (~0.00368ms per sample)