Direct Stream Digital (DSD)
Direct Stream Digital represents audio using a high-rate one-bit stream rather than conventional multibit linear PCM samples.
How can a recording with only one bit per sample represent a quiet voice and a loud orchestra? Direct Stream Digital uses a very fast one-bit sequence whose behavior over time represents the waveform. An individual bit is not a miniature multibit PCM amplitude measurement.
DSD is associated with Super Audio CD and downloadable music. Its usual DSD64 rate is 2,822,400 bits per second per channel: 64 × 44,100.
Pulse density, modulation, and noise shaping
In a simple illustration, map one to +1 and zero to −1. The average of a short bit group then changes with the balance:
| Eight illustrative bits | Ones / zeros | Average of the mapped levels |
|---|---|---|
| 11110000 | 4 / 4 | 0 |
| 11111100 | 6 / 2 | +0.5 |
| 11000000 | 2 / 6 | −0.5 |
This is a teaching model of pulse density, not a DSD encoding algorithm. The actual order of bits and the feedback process matter. Two sequences with the same count in one group can have different high-frequency behavior.
A DSD modulator uses a rapidly running quantizer and feedback. Noise shaping arranges much of the quantization error outside the desired audio band, where subsequent filtering can attenuate it. This lets useful low-frequency amplitude information emerge from a stream whose instantaneous values have only two states.
The very high rate serves the representation
| Common label, 44.1 kHz rate family | One-bit rate per channel |
|---|---|
| DSD64 | 2.8224 MHz |
| DSD128 | 5.6448 MHz |
| DSD256 | 11.2896 MHz |
The rate and one-bit representation belong together. Comparing “2.8 MHz” with “96 kHz PCM” as if they measured one identical precision system discards the most important part of the description.
For raw stereo DSD64, the data rate is 2,822,400 × 1 × 2 = 5.6448 Mbit/s. One minute uses 5,644,800 / 8 × 60 = 42.336 MB before file overhead. This is four times the raw data rate of 44.1 kHz, 16-bit stereo PCM, despite the one-bit samples.
Those arithmetic relationships describe storage demand, not an audible ranking of music releases.
Filtering turns the stream into a usable output
The desired audio and shaped high-frequency error coexist in the sequence. Playback needs suitable reconstruction/filtering so that unwanted high-frequency energy is not simply sent onward as part of the audible signal.
A DSD-to-PCM conversion also filters, then produces multibit values on a new time grid. For an illustrative 88.2 kHz target, DSD64 has 32 input bit positions per output sample interval. A proper converter uses a designed filter and rate conversion; merely counting ones in successive isolated 32-bit groups is not a complete high-quality converter.
The result is a PCM representation of the available audio band. It is not byte-preserving restoration of the original DSD sequence. The filter, target rate, and numerical precision become part of that conversion.
Keep packaging and compression separate
DSF and DSDIFF are containers for DSD-related data. They have different channel-block, bit-packing, and metadata conventions.
DST and WavPack can losslessly compress DSD: after unpacking, the original one-bit sequence can be restored. That operation differs from converting the audio to PCM.
A useful source-to-output trace is therefore:
DSF/DFF reader → any required lossless decompression → DSD sequence → native supported output or filtered PCM conversion → playback.
For a DSD item in Emby, these stages explain why a file remains labeled DSD in the library while the client receives PCM. The stored representation, delivery representation, and DAC input are distinct observations. A DAC can additionally use its own internal processing; “DSD input” describes the accepted signal, not every internal circuit stage.