Click here to Skip to main content
13,629,406 members
Click here to Skip to main content
Add your own
alternative version

Stats

200.7K views
5.3K downloads
81 bookmarked
Posted 25 May 2006
Licenced

Using the G711 standard

, 28 Jul 2006
Rate this:
Please Sign up or sign in to vote.
Implementing the G711 µ-law and a-law codecs in C#.

Introduction

I have been working on a VoIP application, and wanted to implement the G.711 specification, which I found out had two variants: A-law and µ-law. Compounded by the problem of the latter being referred to as "mu-law" and "u-law" in addition to µ-law (ALT-0181, by the way), the documentation for the two is quite poor. Without an ITU login, you can't see the actual standard, and few places online go in to exactly what happens in these encodings. Wikipedia just throws a couple of equations around, but is not terribly helpful. Eventually, I did find various implementations, only one of which was reasonably commented. Of course, that one was the one with the error in it...

So here, in my first CodeProject article, I will explain thoroughly the implementation of G.711 in both of its forms. The code is in C#, but it is simple enough to be ported to, say, Java.

Note that in most contexts, I will use mu instead of µ. Only a few comments and variable names use µ, it's just too strange to think of as a normal character. And although I know the alt-code by heart, it still takes much longer to type.

The Code

The project is arranged into five C# files. One is Program.cs, which is not really important until later, and even then, it isn't so terribly important. The other four are static classes, MuLawEncoder, MuLawDecoder, ALawEncoder, and ALawDecoder, which all do exactly as their names imply.

The static constructors do all of the real work, and store the results in a table. When the Encode or Decode methods are called, they look in the table. At the cost of only a little memory (64 KB per encoder, 0.5KB per decoder), it is definitely worth it.

µ-Law Encoding

The MuLawEncoder class handles the µ-law encoding (surprise!) by looking up values in its own private byte array pcmToMuLawMap. If the index is the unsigned 16-bit PCM value, the value is the unsigned 8-bit µ-law byte.

public const int BIAS = 0x84; //132, or 1000 0100
public const int MAX = 32635; //32767 (max 15-bit integer) minus BIAS

These are the constants used by the encoder. Both will come up later.

static MuLawEncoder()
{
    pcmToMuLawMap = new byte[65536];
    for (int i = short.MinValue; i <= short.MaxValue; i++)
        pcmToMuLawMap[(i & 0xffff)] = encode(i);
}

The static constructor fills the table. It instantiates it as an array of 65536 bytes, and then goes from -37638 to 37637. To make the index of the array not be negative, i is ANDed with 0xffff, making it positive.

The method encode(short i) does exactly what you would expect. However, it is private. This is because the only time it will ever be used is by the constructor.

private static byte encode(int pcm) //16-bit
{
    //Get the sign bit. Shift it for later 
    //use without further modification
    int sign = (pcm & 0x8000) >> 8;
    //If the number is negative, make it 
    //positive (now it's a magnitude)
    if (sign != 0)
        pcm = -pcm;
    //The magnitude must be less than 32635 to avoid overflow
    if (pcm > MAX) pcm = MAX;
    //Add 132 to guarantee a 1 in 
    //the eight bits after the sign bit
    pcm += BIAS;

    /* Finding the "exponent"
    * Bits:
    * 1 2 3 4 5 6 7 8 9 A B C D E F G
    * S 7 6 5 4 3 2 1 0 . . . . . . .
    * We want to find where the first 1 after the sign bit is.
    * We take the corresponding value from
    * the second row as the exponent value.
    * (i.e. if first 1 at position 7 -> exponent = 2) */
    int exponent = 7;
    //Move to the right and decrement exponent until we hit the 1
    for (int expMask = 0x4000; (pcm & expMask) == 0; 
         exponent--, expMask >>= 1) { }

    /* The last part - the "mantissa"
    * We need to take the four bits after the 1 we just found.
    * To get it, we shift 0x0f :
    * 1 2 3 4 5 6 7 8 9 A B C D E F G
    * S 0 0 0 0 0 1 . . . . . . . . . (meaning exponent is 2)
    * . . . . . . . . . . . . 1 1 1 1
    * We shift it 5 times for an exponent of two, meaning
    * we will shift our four bits (exponent + 3) bits.
    * For convenience, we will actually just shift
    * the number, then and with 0x0f. */
    int mantissa = (pcm >> (exponent + 3)) & 0x0f;

    //The mu-law byte bit arrangement 
    //is SEEEMMMM (Sign, Exponent, and Mantissa.)
    byte mulaw = (byte)(sign | exponent << 4 | mantissa);

    //Last is to flip the bits
    return (byte)~mulaw;
}

The comments say everything that needs to be said.

The public methods all access the table, rather than do reprocessing. They are all called MuLawEncode, with different arguments. It is redundant, but so be it. The Encode overloads:

public static byte MuLawEncode(int pcm) { /*...*/ }
public static byte MuLawEncode(short pcm) { /*...*/ }
public static byte[] MuLawEncode(int[] data) { /*...*/ }
public static byte[] MuLawEncode(short[] data) { /*...*/ }
public static void[] MuLawEncode(byte[] data, byte[] target) 
    { /*Suggested by Nathan Allan*/ }
public static byte[] MuLawEncode(byte[] data) {
    int size = data.Length / 2;
    byte[] encoded = new byte[size];
    for (int i = 0; i < size; i++)
    encoded[i] = MuLawEncode((data[2 * i + 1] << 8) | data[2 * i]);
    return encoded;
}

The last takes an array of bytes in Little-Endian order. Thus, it is special, and gets to be displayed.

The last thing in the MuLawEncoder class is the ZeroTrap. Apparently, it is not so great of a thing to send an all-zero µ-law byte, so when the trap is enabled, an all-zero µ-law byte is replaced instead, by 0x02. By default, this trap is disabled.

Normally, the zero trap is a boolean, but here, there is no need. Since the unsigned PCM value 33000 maps to 0x00, we know that if the table reads 0x00, the zero trap is off. See:

public bool ZeroTrap
{
    get { return (pcmToMuLawMap[33000] != 0); }
    set
    {
        byte val = (byte)(value ? 2 : 0);
        for (int i = 32768; i <= 33924; i++)
            pcmToMuLawMap[i] = val;
    }
}

When the zero trap is assigned, the program will go through the table and assign either 0x00 or 0x02 to all of the places that map to 0x00 normally. These are the values in the range [32768, 33924].

µ-Law Decoding

In yet another major surprise, this is done in the MuLawDecoder class. This uses the same table lookup technique as the above, but has an array of type short, since the values are 16-bit signed PCM values.

static MuLawDecoder()
{
    muLawToPcmMap = new short[256];
    for (byte i = 0; i < byte.MaxValue; i++)
        muLawToPcmMap[i] = decode(i);
}

private static short decode(byte mulaw)
{
    //Flip all the bits
    mulaw = (byte)~mulaw;

    //Pull out the value of the sign bit
    int sign = mulaw & 0x80;
    //Pull out and shift over the value of the exponent
    int exponent = (mulaw & 0x70) >> 4;
    //Pull out the four bits of data
    int data = mulaw & 0x0f;

    //Add on the implicit fifth bit (we know 
    //the four data bits followed a one bit)
    data |= 0x10;
    /* Add a 1 to the end of the data by 
    * shifting over and adding one. Why?
    * Mu-law is not a one-to-one function. 
    * There is a range of values that all
    * map to the same mu-law byte. 
    * Adding a one to the end essentially adds a
    * "half byte", which means that 
    * the decoding will return the value in the
    * middle of that range. Otherwise, the mu-law
    * decoding would always be
    * less than the original data. */
    data <<= 1;
    data += 1;
    /* Shift the five bits to where they need
    * to be: left (exponent + 2) places
    * Why (exponent + 2) ?
    * 1 2 3 4 5 6 7 8 9 A B C D E F G
    * . 7 6 5 4 3 2 1 0 . . . . . . . <-- starting bit (based on exponent)
    * . . . . . . . . . . 1 x x x x 1 <-- our data
    * We need to move the one under the value of the exponent,
    * which means it must move (exponent + 2) times
    */
    data <<= exponent + 2;
    //Remember, we added to the original,
    //so we need to subtract from the final
    data -= MuLawEncoder.BIAS;
    //If the sign bit is 0, the number 
    //is positive. Otherwise, negative.
    return (short)(sign == 0 ? data : -data);
}

Again, the comments explain the magic.

And again, the main function is overloaded:

public static short MuLawDecode(byte mulaw) { /*...*/ }
public static short[] MuLawDecode(byte[] data) { /*...*/ }
public static void MuLawDecode(byte[] data, out short[] decoded) { /*...*/ }
public static void MuLawDecode(byte[] data, out byte[] decoded)
{
    int size = data.Length;
    decoded = new byte[size * 2];
    for (int i = 0; i < size; i++)
    {
        //First byte is the less significant byte
        decoded[2 * i] = (byte)(muLawToPcmMap[data[i]] & 0xff);
        //Second byte is the more significant byte
        decoded[2 * i + 1] = (byte)(muLawToPcmMap[data[i]] >> 8);
    }
}

The out parameters are used because otherwise there would be no way to separate the two MuLawDecode functions that both take a byte array. And again, the Little-Endian byte order is displayed.

A-law Encoding

A-law is even worse documented than µ-law. The implementations are even worse in terms of commenting, so this took a bit longer to figure out. In the end, it is indeed similar to µ-law's implementation, despite how different the A-law C code looks from the µ-law C code.

There is no zero trap, and the ALawEncode overloads are identical to the MuLawEncode overloads, so the only difference is the encode(short i) method, and that the MAX is 0x7fff instead of (0x7fff-0x84) like in MuLawEncoder.

private static byte encode(int pcm)
{
    //Get the sign bit. Shift it for later use 
    //without further modification
    int sign = (pcm & 0x8000) >> 8;
    //If the number is negative, 
    //make it positive (now it's a magnitude)
    if (sign != 0)
        pcm = -pcm;
    //The magnitude must fit in 15 bits to avoid overflow
    if (pcm > MAX) pcm = MAX;

    /* Finding the "exponent"
     * Bits:
     * 1 2 3 4 5 6 7 8 9 A B C D E F G
     * S 7 6 5 4 3 2 1 0 0 0 0 0 0 0 0
     * We want to find where the first 1 after the sign bit is.
     * We take the corresponding value 
     * from the second row as the exponent value.
     * (i.e. if first 1 at position 7 -> exponent = 2)
     * The exponent is 0 if the 1 is not found in bits 2 through 8.
     * This means the exponent is 0 even if the "first 1" doesn't exist.
     */
    int exponent = 7;
    //Move to the right and decrement exponent 
    //until we hit the 1 or the exponent hits 0
    for (int expMask = 0x4000; (pcm & expMask) == 0 
         && exponent>0; exponent--, expMask >>= 1) { }

    /* The last part - the "mantissa"
     * We need to take the four bits after the 1 we just found.
     * To get it, we shift 0x0f :
     * 1 2 3 4 5 6 7 8 9 A B C D E F G
     * S 0 0 0 0 0 1 . . . . . . . . . (say that exponent is 2)
     * . . . . . . . . . . . . 1 1 1 1
     * We shift it 5 times for an exponent of two, meaning
     * we will shift our four bits (exponent + 3) bits.
     * For convenience, we will actually just
     * shift the number, then AND with 0x0f. 
     * 
     * NOTE: If the exponent is 0:
     * 1 2 3 4 5 6 7 8 9 A B C D E F G
     * S 0 0 0 0 0 0 0 Z Y X W V U T S (we know nothing about bit 9)
     * . . . . . . . . . . . . 1 1 1 1
     * We want to get ZYXW, which means a shift of 4 instead of 3
     */
    int mantissa = (pcm >> ((exponent == 0) ? 4 : (exponent + 3))) & 0x0f;

    //The a-law byte bit arrangement is SEEEMMMM 
    //(Sign, Exponent, and Mantissa.)
    byte alaw = (byte)(sign | exponent << 4 | mantissa);

    //Last is to flip every other bit, and the sign bit (0xD5 = 1101 0101)
    return (byte)(alaw^0xD5);
}

Even this has only subtle differences. The mask, lack of bias, and the zero exponent weirdness are the key differences.

A-Law Decoding

The only difference between this and MuLawDecoder is the decode(short i) method.

private static short decode(byte alaw)
{
    //Invert every other bit, 
    //and the sign bit (0xD5 = 1101 0101)
    alaw ^= 0xD5;

    //Pull out the value of the sign bit
    int sign = alaw & 0x80;
    //Pull out and shift over the value of the exponent
    int exponent = (alaw & 0x70) >> 4;
    //Pull out the four bits of data
    int data = alaw & 0x0f;

    //Shift the data four bits to the left
    data <<= 4;
    //Add 8 to put the result in the middle 
    //of the range (like adding a half)
    data += 8;
    
    //If the exponent is not 0, then we know the four bits followed a 1,
    //and can thus add this implicit 1 with 0x100.
    if (exponent != 0)
        data += 0x100;
    /* Shift the bits to where they need to be: left (exponent - 1) places
     * Why (exponent - 1) ?
     * 1 2 3 4 5 6 7 8 9 A B C D E F G
     * . 7 6 5 4 3 2 1 . . . . . . . . <-- starting bit (based on exponent)
     * . . . . . . . Z x x x x 1 0 0 0 <-- our data (Z is 0 only when <BR>     * exponent is 0)
     * We need to move the one under the value of the exponent,
     * which means it must move (exponent - 1) times
     * It also means shifting is unnecessary if exponent is 0 or 1.
     */
    if (exponent > 1)
        data <<= (exponent - 1);

    return (short)(sign == 0 ? data : -data);
}

That's it for the encoders and decoders.

Program.cs

The program included in the source package runs the ALawEncoder and MuLawEncoder on a random series of data, and averages the percent errors. It then displays the average errors for each codec for the full range.

Some results:

On the range [1,32767]: µ-Law: 1.14%, A-Law: 1.26%
On the range [-32767,-1]: µ-Law: 1.14%, A-Law: 1.16%

Conclusion

A-law and µ-law are not so complicated, especially when. laid out in plain sight. I hope this is useful. After all, G.711 can make 16-bit samples, take up 8 bits, or 50% compression. At 8KHz sampling, that turns a 128Kbps stream into a 64Kbps stream.

Edited July 28th, 2006 to fix error and add new overload suggested by Nathan Allan

License

This article has no explicit license attached to it but may contain usage terms in the article text or the download files themselves. If in doubt please contact the author via the discussion board below.

A list of licenses authors might use can be found here

Share

About the Author

Marc Sweetgall
United States United States
Marc started proramming on a TI-83 Plus, and slowly moved through the languages: QBasic, VB, C++, and then Java. J# provided a good stepping stone into .NET, and now he uses primarily Java and C#.

You may also be interested in...

Pro
Pro

Comments and Discussions

 
QuestionA-law Decoder/Encoder is not converging? Pin
tassoma8-Jul-18 5:38
membertassoma8-Jul-18 5:38 
QuestionLinear to Mp3 Pin
Member 129343117-Mar-17 0:12
memberMember 129343117-Mar-17 0:12 
QuestionImplement in Net Framework 4.0 or 4.5 Pin
toni firnandes17-Nov-13 23:20
membertoni firnandes17-Nov-13 23:20 
AnswerRe: Implement in Net Framework 4.0 or 4.5 Pin
Parhs18-Jul-14 11:51
memberParhs18-Jul-14 11:51 
QuestionLicensing Under BSD Pin
Member 997809211-Apr-13 10:52
professionalMember 997809211-Apr-13 10:52 
QuestionLots of noise and echo sound Pin
rajesh@198911-Apr-13 0:03
memberrajesh@198911-Apr-13 0:03 
QuestionWav file to ulaw file Pin
zackqwer14-Jan-13 5:54
memberzackqwer14-Jan-13 5:54 
QuestionBug or not? Pin
a.nemec10-Aug-12 0:06
membera.nemec10-Aug-12 0:06 
GeneralPort for Silverlight 4 Pin
Andreas Niedermair2-Jan-12 20:55
memberAndreas Niedermair2-Jan-12 20:55 
QuestionCommercial Licensing for a Port Pin
Member 82061741-Sep-11 13:39
memberMember 82061741-Sep-11 13:39 
QuestionALAW to FLAC Pin
tsdaemon22-Jun-11 6:40
membertsdaemon22-Jun-11 6:40 
QuestionNortel VBK? Pin
Member 167491827-Aug-10 4:41
memberMember 167491827-Aug-10 4:41 
QuestionLicencing Pin
rainer.hochdorfer27-Apr-10 22:08
memberrainer.hochdorfer27-Apr-10 22:08 
GeneralThanks Pin
GSc_Dev28-Oct-08 2:19
memberGSc_Dev28-Oct-08 2:19 
Generalpcm to g.711 Pin
pasalos20-May-08 23:47
memberpasalos20-May-08 23:47 
GeneralPCM to uLaw - lots of noice Pin
Member 158205014-Feb-08 0:45
memberMember 158205014-Feb-08 0:45 
GeneralRe: PCM to uLaw - lots of noice Pin
pasalos28-May-08 2:47
memberpasalos28-May-08 2:47 
Generalported to java Pin
bansal_rajeshkr13-Sep-07 23:17
memberbansal_rajeshkr13-Sep-07 23:17 
GeneralOverhead for memory plz chk Pin
bansal_rajeshkr13-Sep-07 21:51
memberbansal_rajeshkr13-Sep-07 21:51 
GeneralLicensing... for FMJ, an open-source project. Pin
KenLars997-Jun-07 12:28
memberKenLars997-Jun-07 12:28 
AnswerRe: Licensing... for FMJ, an open-source project. Pin
Marc Sweetgall15-Jun-07 19:38
memberMarc Sweetgall15-Jun-07 19:38 
QuestionSource Needed Pin
amessbee22-Feb-07 2:00
memberamessbee22-Feb-07 2:00 
AnswerRe: Source Needed Pin
Marc Sweetgall15-Jun-07 19:41
memberMarc Sweetgall15-Jun-07 19:41 
QuestionLicensing? Pin
neilcam17-Dec-06 23:31
memberneilcam17-Dec-06 23:31 
QuestionTrim silence from recording Pin
koenvi6-Sep-06 22:37
memberkoenvi6-Sep-06 22:37 

General General    News News    Suggestion Suggestion    Question Question    Bug Bug    Answer Answer    Joke Joke    Praise Praise    Rant Rant    Admin Admin   

Use Ctrl+Left/Right to switch messages, Ctrl+Up/Down to switch threads, Ctrl+Shift+Left/Right to switch pages.

Permalink | Advertise | Privacy | Cookies | Terms of Use | Mobile
Web01 | 2.8.180712.1 | Last Updated 28 Jul 2006
Article Copyright 2006 by Marc Sweetgall
Everything else Copyright © CodeProject, 1999-2018
Layout: fixed | fluid