Decoding Sony PlayStation VAG Audio Streams into Universal 16-Bit WAV

When the original Sony PlayStation launched in 1994, its custom Sound Processing Unit (SPU) delivered CD-quality synthesized audio, digital reverberation, and 24 hardware audio voices. To pack thousands of dialogue clips, sword swings, and footsteps onto CD-ROMs, Sony developed the VAG audio format (.vag, .vagp).

From Final Fantasy VII and Metal Gear Solid to Crash Bandicoot and Castlevania: Symphony of the Night, VAG was the voice of a gaming generation. Here is how Sony SPU-ADPCM compression functions and how convrtr decodes it in your browser.

The VAG Header

A standard VAG file begins with a 48-byte or 64-byte big-endian header:

• Bytes 0..3 (Magic): ASCII 'VAGp' (0x56 0x41 0x47 0x70), or rarely 'pGAV' on little-endian targets.
• Bytes 4..7 (Version): Typically 0x00000003 (VAG v3) or 0x00000020 (VAG v2).
• Bytes 12..15 (Data Size): Total byte length of the ADPCM payload.
• Bytes 16..19 (Sample Rate): Playback frequency in Hertz (typically 22,050 Hz, 32,000 Hz, or 44,100 Hz).
• Bytes 32..47 (Name): Up to 16 ASCII characters identifying the sound cue (e.g. SLASH_01).

If a 64-byte padded header is used, bytes 48 through 63 are filled with zeros before the audio blocks begin.

SPU-ADPCM 16-Byte Block Format

Sony's hardware ADPCM compression reduces 16-bit linear audio to 4 bits per sample—a 4:1 reduction in storage footprint. Audio is structured in uniform 16-byte blocks, each expanding into 28 linear PCM samples:

• Byte 0 (Predictor & Shift):
  • Lower nibble (byte & 0x0f): Shift factor (0 through 12).
  • Upper nibble ((byte >> 4) & 0x07): Filter coefficient index (0 through 4).
• Byte 1 (Flags): Loop flags including Loop End (0x01), Loop Repeat (0x02), and Loop Start (0x04).
• Bytes 2..15 (14 bytes): 28 4-bit signed samples packed in low-nibble-first order.

The 2-Pole SPU Linear Prediction Filter

To reconstruct the acoustic waveform, the hardware applies a 2-pole Infinite Impulse Response (IIR) filter using two previous sample values (s1 and s2) and fixed coefficient pairs:

| Filter Index | Positive Coeff (K1) | Negative Coeff (K2) | |---|---|---| | 0 | 0.0 | 0.0 | | 1 | 60 / 64 (0.9375) | 0.0 | | 2 | 115 / 64 (1.7968) | -52 / 64 (-0.8125) | | 3 | 98 / 64 (1.5312) | -55 / 64 (-0.8593) | | 4 | 122 / 64 (1.9062) | -60 / 64 (-0.9375) |

For every 4-bit nibble d, the sample is sign-extended and decoded as:

raw = d << (12 - shift)
predicted = raw + (s1 * K1 + s2 * K2 + 32) / 64

The predicted sample is clamped to the 16-bit signed integer range (-32,768 to 32,767) and stored in the sample buffer.

Universal RIFF WAV Synthesis

Once all SPU blocks are decoded into signed 16-bit PCM samples, convrtr synthesizes a canonical 44-byte RIFF WAVE header with the native sample rate and little-endian sample stream.

The resulting .wav file plays natively in any web browser, media player, or digital audio workstation with zero external software required and complete confidentiality.