33
\$\begingroup\$

Task

Implement a program in minimum bytes of source or binary code that does voice recognition of a voice sample (me saying "yes", "yeah" or "no" in voice or in whisper, plainly or quirkily) based on training samples with maximum accuracy.

The program should read train/yes0.wav, train/no0.wav, train/yes1.wav and so on (there are 400 yeses and 400 noes in training dataset), then start reading inputs/0.wav, inputs/1.wav until it fails to find the file, analysing it and outputting "yes" or "no" (or other word for intermediate answer).

If you want, you may pre-train the program instead of reading train/, but the resulting data table counts towards the score (and beware of overfitting to training samples - they don't overlap with examination ones). Better to include the program used to produce the data table as an addendum in this case.

All sample files are little endian 16-bit stereo WAV files, just from laptop mic, without filtering/noise reduction.

Limits

Forbidden features:

  • Using network;
  • Trying to reach the answers file inputs/key;
  • Subverting the runner program that calculates accuracy;
  • Using existing recognition libraries. Linking to FFT implementation not allowed: only external math functions taking constant amount of information (like sin or atan2) are allowed; If you want FFT, just add it's implementation to your program source code (it can be multilingual if needed).

Resource limits:

  • The program should not take more than 30 minutes of CPU time on my i5 laptop. If it takes more, only output produced in the first 30 minutes are counted and eveything else is assumed a half-match;
  • Memory limit: 1GB (including any temporary files);

Tools

The tools/runner program automatically runs your solution and calculates the accuracy.

$ tools/runner solutions/example train train/key 
Accuracy: 548 ‰

It can examine the program using training data or using actual exam data. I'm going to try submitted answers on examination dataset and publish results (accuracy percentage) until I make the dataset public.

Scoring

There are 5 classes of solution depending on accuracy:

  • All samples guessed correctly: Class 0;
  • Accuracy 950-999: Class 1;
  • Accuracy 835-950: Class 2;
  • Accuracy 720-834: Class 3;
  • Accuracy 615-719: Class 4;

Inside each class, the score is number of bytes the solution takes.

Accepted answer: smallest solution in the best nonempty class.

Links

All samples should be considered CC-0 (Public Domain), scripts and programs should be considered MIT.

Example solution

It provides very poor quality of recognition, just shows how to read files and output answers

#define _BSD_SOURCE
#include <stdio.h>
#include <assert.h>
#include <endian.h>


#define Nvols 30

#define BASS_WINDOW 60
#define MID_WINDOW 4

struct training_info {
    double bass_volumes[Nvols];
    double mid_volumes[Nvols];
    double treble_volumes[Nvols];
    int n;
};


struct training_info yes;
struct training_info no;

static int __attribute__((const)) mod(int n, int d) {
    int m = n % d;
    if (m < 0) m+=d;
    return m;
}

// harccoded to 2 channel s16le
int get_file_info(const char* name, struct training_info *inf) {
    FILE* in = fopen(name, "rb");

    if (!in) return -1;

    setvbuf(in, NULL, _IOFBF, 65536);

    inf->n = 1;

    fseek(in, 0, SEEK_END);
    long filesize = ftell(in);
    fseek(in, 128, SEEK_SET);
    filesize -= 128; // exclude header and some initial samples

    int nsamples = filesize / 4; 

    double bass_a=0, mid_a=0;
    const int HISTSIZE  = 101;
    double xhistory[HISTSIZE];
    int histpointer=0;
    int histsize = 0;

    //FILE* out = fopen("debug.raw", "wb");

    int i;
    for (i=0; i<Nvols; ++i) {
        int j;

        double total_vol = 0;
        double bass_vol = 0;
        double mid_vol = 0;
        double treble_vol = 0;

        for (j=0; j<nsamples / Nvols; ++j) {
            signed short int l, r; // a sample
            if(fread(&l, 2, 1, in)!=1) break;
            if(fread(&r, 2, 1, in)!=1) break;
            double x = 1/65536.0 * ( le16toh(l) + le16toh(r) );


            bass_a += x;
            mid_a  += x;


            if (histsize == HISTSIZE-1) bass_a   -= xhistory[mod(histpointer-BASS_WINDOW,HISTSIZE)];
            if (histsize == HISTSIZE-1) mid_a    -= xhistory[mod(histpointer-MID_WINDOW ,HISTSIZE)];

            double bass = bass_a / BASS_WINDOW;
            double mid = mid_a / MID_WINDOW - bass;
            double treble = x - mid_a/MID_WINDOW;

            xhistory[histpointer++] = x;
            if(histpointer>=HISTSIZE) histpointer=0;
            if(histsize < HISTSIZE-1) ++histsize;

            total_vol  += bass*bass + mid*mid + treble*treble;
            bass_vol   += bass*bass;
            mid_vol    += mid*mid;
            treble_vol += treble*treble;


            /*
            signed short int y;
            y = 65536 * bass;

            y = htole16(y);
            fwrite(&y, 2, 1, out);
            fwrite(&y, 2, 1, out);
            */
        }

        inf->bass_volumes[i] = bass_vol / total_vol;
        inf->mid_volumes[i] = mid_vol / total_vol;
        inf->treble_volumes[i] = treble_vol / total_vol;

        //fprintf(stderr, "%lf %lf %lf    %s\n", inf->bass_volumes[i], inf->mid_volumes[i], inf->treble_volumes[i], name);
    }
    fclose(in);

    return 0;
}

static void zerotrdata(struct training_info *inf) {
    int i;
    inf->n = 0;
    for (i=0; i<Nvols; ++i) {
        inf->bass_volumes[i] = 0;
        inf->mid_volumes[i] = 0;
        inf->treble_volumes[i] = 0;
    }
}

static void train1(const char* prefix, struct training_info *inf) 
{
    char buf[50];

    int i;

    for(i=0;; ++i) {
        sprintf(buf, "%s%d.wav", prefix, i);
        struct training_info ti;
        if(get_file_info(buf, &ti)) break;

        ++inf->n;

        int j;
        for (j=0; j<Nvols; ++j) {
            inf->bass_volumes[j]   += ti.bass_volumes[j];
            inf->mid_volumes[j]    += ti.mid_volumes[j];
            inf->treble_volumes[j] += ti.treble_volumes[j];
        }
    }

    int j;
    for (j=0; j<Nvols; ++j) {
        inf->bass_volumes[j]   /= inf->n;
        inf->mid_volumes[j]    /= inf->n;
        inf->treble_volumes[j] /= inf->n;
    }
}

static void print_part(struct training_info *inf, FILE* f) {
    fprintf(f, "%d\n", inf->n);
    int j;
    for (j=0; j<Nvols; ++j) {
        fprintf(f, "%lf %lf %lf\n", inf->bass_volumes[j], inf->mid_volumes[j], inf->treble_volumes[j]);
    }
}

static void train() {
    zerotrdata(&yes);
    zerotrdata(&no);

    fprintf(stderr, "Training...\n");

    train1("train/yes", &yes);
    train1("train/no", &no);

    fprintf(stderr, "Training completed.\n");

    //print_part(&yes, stderr);
    //print_part(&no, stderr);

    int j;
    for (j=0; j<Nvols; ++j) {
        if (yes.bass_volumes[j]   > no.bass_volumes[j]) {   yes.bass_volumes[j] = 1;   no.bass_volumes[j] = 0; }
        if (yes.mid_volumes[j]    > no.mid_volumes[j]) {    yes.mid_volumes[j] = 1;    no.mid_volumes[j] = 0; }
        if (yes.treble_volumes[j] > no.treble_volumes[j]) { yes.treble_volumes[j] = 1; no.treble_volumes[j] = 0; }
    }
}


double delta(struct training_info *t1, struct training_info *t2) {
    int j;
    double d = 0;
    for (j=0; j<Nvols; ++j) {
        double rb = t1->bass_volumes[j] - t2->bass_volumes[j];
        double rm = t1->mid_volumes[j] - t2->mid_volumes[j];
        double rt = t1->treble_volumes[j] - t2->treble_volumes[j];
        d += rb*rb + rm*rm + rt*rt;
    }
    return d;
}

int main(int argc, char* argv[])
{
    (void)argc; (void)argv;

    train();


    int i;

    int yes_count = 0;
    int no_count = 0;

    for (i=0;; ++i) {
        char buf[60];
        sprintf(buf, "inputs/%d.wav", i);

        struct training_info ti;

        if(get_file_info(buf, &ti)) break;

        double dyes = delta(&yes, &ti);
        double dno = delta(&no, &ti);

        //printf("%lf %lf %s ", dyes, dno, buf);

        if (dyes > dno) {
            printf("no\n");
            ++no_count;
        } else  {
            printf("yes\n");
            ++yes_count;
        }
    }

    fprintf(stderr, "yeses: %d noes: %d\n", yes_count, no_count);

}
\$\endgroup\$
  • 5
    \$\begingroup\$ no fft libraries? Why? \$\endgroup\$ – John Dvorak Sep 1 '14 at 12:02
  • 1
    \$\begingroup\$ What about built-in FFT functions? What exactly counts as external? Also, what counts as a math library function? Are we allowed to use sum, or we need to use foldl (+) 0 (foldl not being math-specific, and + not being variadic)? \$\endgroup\$ – John Dvorak Sep 1 '14 at 12:14
  • 1
    \$\begingroup\$ still... you're effectively forbidding sum. I guess that's not your intention? \$\endgroup\$ – John Dvorak Sep 1 '14 at 12:17
  • 1
    \$\begingroup\$ What's the exact definition of math functions? Those that are specialised to operate on numbers? What about a generic "sum" function that uses addition for numbers, but concatenation for strings? Is this sum now allowed? \$\endgroup\$ – John Dvorak Sep 1 '14 at 12:18
  • 1
    \$\begingroup\$ What about J's vector operations? Are they disallowed? \$\endgroup\$ – John Dvorak Sep 1 '14 at 12:20
27
\$\begingroup\$

C++11 (gcc; 1639 1625 1635 bytes, Class 1, score=983, 960)

Let's get it started. It's probably the longest code I've ever shortened...

#include <bits/stdc++.h>
#define $ complex<double>
#define C vector<$>
#define I int
#define D double
#define P pair<D,I>
#define Q pair<D,D>
#define E vector<D>
#define V vector<P>
#define W vector<Q>
#define S char*
#define Z(x)if(fread(&x,2,1,y)!=1)break;
#define B push_back
#define F(i,f,t)for(I i=f;i<t;i++)
#define _ return
#define J first
#define L second
const I K=75,O=16384;using namespace std;C R(C&i,I s,I o=0,I d=1){if(!s)_ C(1,i[o]);C l=R(i,s/2,o,d*2),h=R(i,s/2,o+d,d*2);C r(s*2);$ b=exp($(0,-M_PI/s)),f=1;F(k,0,s){r[k]=l[k]+f*h[k];r[k+s]=l[k]-f*h[k];f*=b;}_ r;}C T(C&i){_ R(i,i.size()/2);}char b[O];E U(S m){FILE*y;if(!(y=fopen(m,"r")))_ E();setvbuf(y,b,0,O);fseek(y,0,2);I z=ftell(y)/4-32;fseek(y,128,0);C p;F(i,0,z){short l,r;Z(l);Z(r);if(i&1)p.B($(0.5/O*le16toh(l),0));}p.resize(O);E h(O),n(O);p=T(p);F(j,0,O)h[j]=norm(p[j])/O;F(i,1,O-1)n[i]=(h[i-1]+h[i+1]+h[i]*8)/10;fclose(y);_ n;}W G(E&d){V p;F(i,3,O/2-3)if(d[i]==*max_element(d.begin()+i-3,d.begin()+i+4))p.B({d[i],i});sort(p.begin(),p.end(),greater<P>());W r;F(i,3,K+3)r.B({D(p[i].L)/O*22050,log(p[i].J)+10});D s=0;F(i,0,K)s+=r[i].L;F(i,0,K)r[i].L/=s;_ r;}char m[O];I X(S p,W&r){E t(O),h(t);I f=0;while(1){sprintf(m,"%s%d.wav",p,f++);h=U(m);if(!h.size())break;F(j,0,O)t[j]+=h[j];}F(j,0,O)t[j]/=f;r=G(t);}D Y(W&a,W&b){D r=0;F(i,0,K){D d=b[i].L;F(j,0,K)if(abs((b[i].J-a[j].J)/b[i].J)<0.015)d=min(d,abs(b[i].L-a[j].L));r+=d;}_ r;}I H(S p,W&y,W&n){I f=0;while(1){sprintf(m,"%s%d.wav",p,f++);E h=U(m);if(!h.size())break;W p=G(h);D q=Y(p,y),r=Y(p,n);printf(abs(q-r)<=0.01?"?\n":q<r?"yes\n":"no\n");}}I main(){W y,n;X("train/yes",y);X("train/no",n);H("inputs/",y,n);}

"Ungolfed" (although it's hard to call an over 1.5K source code golfed):

#include <iostream>
#include <stdio.h>
#include <string>
#include <algorithm>
#include <vector>
#include <math.h>
#include <complex>
#include <endian.h>
#include <functional>

using namespace std;

typedef complex<double> CD;

vector<CD> run_fft(vector<CD>& input, int offset, int size, int dist){
    if(size == 1){
        return vector<CD>(1, input[offset]);
    }
    vector<CD> partLow = run_fft(input, offset, size/2, dist*2),
               partHi  = run_fft(input, offset+dist, size/2, dist*2);

    vector<CD> result(size);
    CD factorBase = exp(CD(0, (inv?2:-2)*M_PI/size)), factor = 1;

    for(int k = 0; k < size/2; k++){
        result[k] = partLow[k] + factor*partHi[k];
        result[k+size/2] = partLow[k] - factor*partHi[k];
        factor *= factorBase;
    }
    return result;
}

vector<CD> fft(vector<CD>& input){
    int N = input.size();
    return run_fft(input, 0, N, 1);
}



const int MAX_BUF = 65536;
const int PWR_TWO = 16384;
const int NUM_CHECK = 75;
int sampling;

char buf[MAX_BUF];
vector<double> read_data(char* filenam){
    FILE* fp = fopen(filenam, "r");
    if(!fp)
        return vector<double>();
    setvbuf(fp, buf, _IOFBF, MAX_BUF);

    fseek(fp, 0, SEEK_END);
    int filesiz = ftell(fp);
    fseek(fp, 128, SEEK_SET);
    filesiz -= 128;

    int insamp = filesiz / 4;
    int freqsamp = 2,
        act_mod = 0;
    sampling = 44100 / freqsamp;
    int inputSize;

    vector<CD> input;

    for(int i = 0; i < insamp; i++){
        signed short int l, r;
        if(fread(&l, 2, 1, fp) != 1) break;
        if(fread(&r, 2, 1, fp) != 1) break;

        double act = 1/32768.0 * (le16toh(l));

        if((++act_mod) == freqsamp){
            inputSize++;
            input.push_back(CD(act,0));
            act_mod = 0;
        }
    }
    inputSize = input.size();

    //printf("%s\n", filenam);
    int numParts = (inputSize+PWR_TWO-1)/PWR_TWO;
    double partDelta = (double)inputSize / numParts, actDelta = 0;
    vector<CD> ndata(PWR_TWO);
    for(int i = 0; i < numParts; i++){
        vector<CD> partInput(PWR_TWO);
        int from = floor(actDelta),
            to = floor(actDelta)+PWR_TWO;

        for(int j = from; j < to; j++)
            partInput[j-from] = input[j];

        vector<CD> partData = fft(partInput);
        for(int j = 0; j < PWR_TWO; j++)
            ndata[j] += partData[j]*(1.0/numParts);
    }


    vector<double> height(PWR_TWO);
    for(int i = 0; i < PWR_TWO; i++)
        height[i] = norm(ndata[i])/PWR_TWO;

    vector<double> nheight(height);
    nheight[0] = (height[0]*0.8 + height[1]*0.1)/0.9;
    nheight[PWR_TWO-1] = (height[PWR_TWO]*0.8 + height[PWR_TWO-1]*0.1)/0.9;
    for(int i = 1; i < PWR_TWO-1; i++)
        nheight[i] = height[i-1]*0.1 + height[i]*0.8 + height[i+1]*0.1;

    fclose(fp);

    return nheight;
}


vector< pair<double,double> > get_highest_peaks(vector<double>& freqData){
    vector< pair<double,int> > peaks;

    for(int i = 3; i < PWR_TWO/2-3; i++){
        if(freqData[i] == *max_element(freqData.begin()+i-3, freqData.begin()+i+4)){
            peaks.push_back(make_pair(freqData[i], i));
        }
    }

    sort(peaks.begin(), peaks.end(), greater< pair<double,int> >());

    vector< pair<double,double> > res;
    for(int i = 3; i < NUM_CHECK+3; i++){
        res.push_back(make_pair((double)(peaks[i].second)/PWR_TWO*sampling, log(peaks[i].first)+10));
    }

    double sum_res = 0;
    for(int i = 0; i < NUM_CHECK; i++)
        sum_res += res[i].second;
    for(int i = 0; i < NUM_CHECK; i++)
        res[i].second /= sum_res;

    /*for(int i = 0; i < NUM_CHECK; i++)
        printf("%12lf %12lf\n", res[i].first, res[i].second);
    printf("\n");*/

    return res;
}


void train(char* dir, const char* type, vector< pair<double,double> >& res){
    vector<double> result(PWR_TWO), height(PWR_TWO);

    int numFile = 0;
    while(true){
        char filenam[256];
        snprintf(filenam, 255, "%s/%s%d.wav", dir, type, numFile);
        height = read_data(filenam);

        if(height.size() == 0)
            break;

        for(int j = 0; j < PWR_TWO; j++)
            result[j] += height[j];

        numFile++;
    }
    fprintf(stderr, "trained %s on %d files\n", type, numFile);

    for(int j = 0; j < PWR_TWO; j++)
        result[j] /= numFile;

    res = get_highest_peaks(result);
}


double dist_ab(vector< pair<double,double> >& A, vector< pair<double,double> >& B){
    double result = 0;
    for(int i = 0; i < NUM_CHECK; i++){
        double add = B[i].second;

        for(int j = 0; j < NUM_CHECK; j++){
            double dist = (B[i].first-A[j].first)/B[i].first;
            if(abs(dist) < 0.015)
                add = min(add, abs(B[i].second - A[j].second));
        }
        result += add;
    }
    return result;
}


void trial(char* dir, const char* pref, vector< pair<double,double> >& yes,
                                        vector< pair<double,double> >& no){
    int numFile = 0;
    int numYes = 0, numDunno = 0, numNo = 0;
    while(true){
        char filenam[256];
        snprintf(filenam, 255, "%s/%s%d.wav", dir, pref, numFile);

        vector<double> height = read_data(filenam);
        if(height.size() == 0)
            break;

        vector< pair<double,double> > peaks = get_highest_peaks(height);


        double distYes = dist_ab(peaks, yes),
               distNo = dist_ab(peaks, no);

        if(abs(distYes-distNo) <= 0.01){
            printf("dunno\n");
            numDunno++;
        } else if(distYes < distNo){
            printf("yes\n");
            numYes++;
        } else {
            printf("no\n");
            numNo++;
        }
        //printf(" (%lf %lf)\n", distYes, distNo);

        numFile++;
    }
}


int main(int argc, char** argv){
    vector< pair<double,double> > yes, no;


    train("train", "yes", yes);
    train("train", "no", no);

    trial("inputs", "", yes, no);
}

I have no freaking idea if it will work correctly on real data set (I bet it won't, but I must have a try).

How it works:

  1. Take N=214 samples from left channel, each in equal time span. Normalize them so that min value=0 and max value=1.
  2. Process them using FFT. Now we switched from time domain to frequency domain. We might say that 0th cell of resulting array is 0Hz equivalent and 213-1st cell is 22050Hz equivalent (that is because I took every other sample from L channel, so my sampling is 22050Hz instead of WAV's frequency, 44100Hz).
  3. Find the mean of all such signals - call it "mean frequency distribution". Find K highest peaks in such distribution (here K=75), omitting first few (probably noise) and find their strength. I used log(mean distribution)+10 and then normalized so that the sum of the biggest peaks was 1.
  4. We have two "peak distributions" - one for Yes, second for No. If we have a WAV to test, we transform it in the same way as before (steps 1, 2, 3) and get distribution D. Then we have to check which distribution (Y/N) D is more similar to. I used the following approach: for each peak in Y/N, try to find it in D. If we find it (approximately), the score for this peak is absolute difference between Y/N's and D's strength; in opposite case, it's Y/N's strength (we assume it's always positive). Better (smaller) score wins. If the results are very close (I used 0.01 absolute difference), output dunno.

As I said, probably on final tests it will be classified as "even worse than random". Of course, I hope it won't :D

Edit: fixed bug (forgot to close the files).

\$\endgroup\$
  • 1
    \$\begingroup\$ You are in luck if it performs worse than random. You literally only need to change one byte - distYes > distNo, and it will do better than random. Or to put it another way, it would be quite astonishing if you could guess the outcome of a coin flip incorrectly 100 times in a row! And it's not unheard of that simple algorithms out-perform than more complex ones, so +1 and I wish you good luck. \$\endgroup\$ – blutorange Sep 2 '14 at 16:06
  • \$\begingroup\$ Testing... It terminates prematurely due to EMFILE (Too many open files)... Trying to fix... \$\endgroup\$ – Vi. Sep 2 '14 at 16:39
  • \$\begingroup\$ Bumped maximum open file counter, now it works. Results: training dataset: Accuracy: 983 ‰; Time: 0m27.570s;; examination dataset: Accuracy: 960 ‰; Time: 0m32.957s. Good job. \$\endgroup\$ – Vi. Sep 2 '14 at 16:44
  • \$\begingroup\$ Okay, I fixed this. 10 bytes more. :) \$\endgroup\$ – mnbvmar Sep 2 '14 at 17:56
  • \$\begingroup\$ Incredible use of #defines :P \$\endgroup\$ – qwr Sep 6 '14 at 22:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.