Tag: Delta-Sigma Modulation

While multi-bit Linear Pulse Code Modulation (LPCM) represents audio by sampling at moderate rates ($44.1\text{ kHz}$ to $192\text{ kHz}$) using wide 16-bit or 24-bit binary words, Direct Stream Digital (DSD) inverts this paradigm completely. Co-developed by Sony and Philips in 1999 as the underlying storage architecture for the Super Audio CD (SACD) and long-term archival of legacy analog master tapes, DSD encodes continuous analog waveforms using a 1-bit word length (a single binary digit, 0 or 1, representing $-1$ or $+1$) clocked at an extreme ultrasonic sampling rate starting at $2.8224\text{ MHz}$ ($64$ times the $44.1\text{ kHz}$ CD sampling rate).

How can a 1-bit quantizer—which possesses an intrinsic raw Signal-to-Quantization-Noise Ratio of barely $7.78\text{ dB}$ ($6.02 \times 1 + 1.76\text{ dB}$)—deliver a pristine, audiophile dynamic range exceeding $120\text{ dB}$ across the $20\text{ Hz}$ to $20\text{ kHz}$ audible band? The answer lies in the brilliant control-loop mathematics of High-Order Delta-Sigma ($\Delta\Sigma$) Modulation and aggressive ultrasonic noise shaping. This comprehensive engineering reference explores Pulse Density Modulation (PDM), 5th-to-7th-order loop filter stability, DSD64 through DSD1024 specifications, DoP (DSD over PCM) transport framing, ultrasonic amplifier intermodulation hazards, and Digital eXtreme Definition (DXD) studio production workflows.

In a 1-bit DSD stream (formally known in electrical engineering as Pulse Density Modulation, PDM), there are no multi-bit numbers representing instantaneous voltage levels. Instead, the instantaneous analog voltage amplitude at any moment in time is represented by the local spatial density of 1 pulses versus 0 pulses inside a tiny sliding time window:

• Maximum Positive Peak ($+V_{max}$): Represented by a dense, continuous train of ones (11111111...).
• Maximum Negative Trough ($-V_{max}$): Represented by a dense, continuous train of zeros (00000000...).
• Zero-Crossing Silence ($0.0\text{ V}$): Represented by an alternating 50% duty-cycle pattern (10101010... or randomized dithered equivalents) whose average DC value over a low-pass filter window integrates strictly to zero.

In theory, converting a 1-bit PDM stream back into a continuous analog waveform requires nothing more than passing the high-speed 1-bit switch output through a simple analog RC/LC low-pass filter that averages the pulse density over time, removing the megahertz switching carrier and leaving the pure baseband audio wave behind.

If you simply sample an audio signal at $2.8224\text{ MHz}$ ($64\times$ oversampling) with a basic 1-bit comparator, the $64\times$ oversampling spreads the 1-bit quantization noise power across a $1.4112\text{ MHz}$ Nyquist bandwidth instead of $22.05\text{ kHz}$. Since every doubling of the sampling rate reduces the in-band ($0\text{–}20\text{ kHz}$) noise floor by $3.01\text{ dB}$ ($0.5\text{ bits}$), $64\times$ oversampling ($2^6$) alone only improves the $7.78\text{ dB}$ SNR by $6 \times 3.01 = 18.06\text{ dB}$, yielding a dismal $25.8\text{ dB}$ SNR—nowhere near high-fidelity audio.

To transform $25.8\text{ dB}$ into $120\text{ dB}$, the 1-bit quantizer is embedded inside a negative feedback control loop containing a cascade of $L$ discrete-time integrators—forming an $L$-th Order Delta-Sigma ($\Delta\Sigma$) Modulator.

Let X(z) = Input Signal, E(z) = Quantization Error, H(z) = Loop Integrator Filter:

Modulator Output Y(z) = STF(z) * X(z) + NTF(z) * E(z)

1. Signal Transfer Function (STF):
   STF(z) = H(z) / (1 + H(z))   --> Approaches 1.0 (Unity Gain Low-Pass) in the 0–20 kHz audio band

2. Noise Transfer Function (NTF):
   NTF(z) = 1 / (1 + H(z))      --> Approaches (1 - z^-1)^L (High-Pass Filter) in the 0–20 kHz band

3. Theoretical In-Band Dynamic Range for L-th Order Modulator at Oversampling Ratio (OSR):
   SNR_max (dB) = 6.02 * N + 1.76 - 5.17 * L + 10 * (2*L + 1) * log10(OSR)

Examine how the loop filter $H(z)$ acts differently on the music signal versus the quantization noise: at low audible frequencies ($20\text{ Hz – }20\text{ kHz}$), the integrator gain $H(z)$ is enormous ($\rightarrow \infty$). Consequently, the Signal Transfer Function $\frac{H(z)}{1 + H(z)}$ equals $1.0$ (passing the music untouched), while the Noise Transfer Function $\frac{1}{1 + H(z)}$ approaches $0.0$ (aggressively suppressing quantization noise in the audible band and shoving that conserved noise energy upward into the ultrasonic spectrum between $50\text{ kHz}$ and $1.4\text{ MHz}$)!

In a 1st-order or 2nd-order Delta-Sigma modulator, the feedback loop is unconditionally stable, but the noise shaping slope is too shallow to reach $120\text{ dB}$ SNR at $2.8224\text{ MHz}$. Consequently, commercial DSD encoders (such as the Sony Sonoma or Philips Grimm ADCs) employ 5th-order to 7th-order loop filters with resonator zeroes placed inside the audible band to dig a deep $120\text{ dB}$ noise trench from $20\text{ Hz}$ to $20\text{ kHz}$.

However, high-order ($L \ge 3$) 1-bit feedback loops are only conditionally stable. Governed by Lee’s Rule of Stability, if the peak input signal modulation index exceeds roughly $50\%$ of the 1-bit quantizer’s full-scale rails ($\approx -6\text{ dB}$ relative to an all-ones bitstream), the internal integrators saturate and the loop breaks into violent, unrecoverable oscillation. Therefore, the Scarlet Book SACD standard defines $0\text{ dB SACD}$ as a $50\%$ modulation density index, reserving the upper $3.1\text{ dB}$ of modulation headroom strictly to prevent modulator overload.

While the original 1999 SACD optical disc specification locked the DSD clock to $64 \times 44.1\text{ kHz} = 2.8224\text{ MHz}$ (now termed DSD64), modern solid-state storage and high-speed USB/I2S DAC interfaces have enabled vastly higher 1-bit oversampling multiples. Doubling the DSD clock rate pushes the ultrasonic noise-shaping mound an entire octave higher in frequency, relaxing analog low-pass reconstruction filter requirements and virtually eliminating ultrasonic noise intrusion near $20\text{ kHz}$:

A critical engineering nuance of DSD64 is what happens immediately above $20\text{ kHz}$. Because a 5th-order modulator pushes all the quantization noise out of the audible band, the noise floor rises at a steep $30\text{ dB/octave}$ slope starting at $21\text{ kHz}$. By $80\text{ kHz}$ to $100\text{ kHz}$, the integrated ultrasonic quantization noise reaches $-20\text{ dBFS}$ to $-15\text{ dBFS}$!

Here lies the central mathematical paradox of 1-bit DSD production: you cannot perform basic digital signal processing (such as a $+0.5\text{ dB}$ volume change, mixing two tracks together, or applying a Parametric EQ filter) directly on a 1-bit stream without immediately destroying its 1-bit structure!

Consider why: multiplying a 1-bit stream (-1 or +1) by a volume coefficient of 0.85 (a 16-bit number) immediately produces a 17-bit multi-bit output word. Similarly, an Infinite Impulse Response (IIR) biquad equalizer filter requires recursive multiplication and accumulation across 32-bit or 64-bit internal registers. To output a 1-bit DSD stream again after a single EQ band, you would have to pass that 64-bit intermediate signal through a brand-new 5th-order 1-bit Delta-Sigma modulator—stacking a second layer of ultrasonic noise shaping on top of the first!

When Merging Technologies developed the Pyramix workstation for Sony and Philips to edit and master SACD recordings in the early 2000s, they needed a way to crossfade, equalize, and mix DSD tracks without cascading multiple 1-bit modulators. Their solution was DXD (Digital eXtreme Definition)—which is simply 24-bit (or 32-bit float) Linear PCM running at $352.8\text{ kHz}$ ($8 \times 44.1\text{ kHz}$, or exactly $\frac{1}{8}$th of the DSD64 clock rate).

When a DSD64 recording is imported into a DXD mastering session, a gentle decimation filter converts the 1-bit $2.8224\text{ MHz}$ stream into 32-bit floating-point samples at $352.8\text{ kHz}$ (whose $176.4\text{ kHz}$ Nyquist limit preserves all the ultrasonic transient speed of DSD while stripping the upper megahertz noise). Engineers perform all mixing, EQ, and dynamics processing in distortion-free 32-bit/64-bit floating-point DXD, and then pass the finished master through a single final Delta-Sigma modulator to render the release DSD64/DSD256 master.

In consumer and studio playback systems, DSD audio is stored and transmitted using specific container formats and framing protocols:

• DSDIFF (.dff): The original studio interchange file format created by Philips for SACD authoring. While compact, .dff lacks standardized support for embedded ID3v2 metadata tags and album artwork.
• Sony DSF (.dsf): The modern universal standard for DSD files. It wraps raw 1-bit PDM audio blocks inside a standard container that includes full ID3v2 tag metadata and album art support.
• WavPack DSD (.wv): Losslessly compresses .dsf and .dff files by $35\%$ to $55\%$ using specialized 1-bit entropy prediction with $100\%$ bit-for-bit reconstruction upon playback.
• DoP (DSD over PCM Open Standard): Early operating systems (macOS CoreAudio, Windows, Linux ALSA) and S/PDIF/AES3 digital cables only knew how to transport 24-bit PCM samples, not raw 1-bit DSD streams. Created by dCS, Playback Designs, and Merging Technologies, DoP packs 16 consecutive 1-bit DSD samples into the lower 16 bits of a standard 24-bit / $176.4\text{ kHz}$ PCM word ($16 \times 176.4\text{ kHz} = 2.8224\text{ MHz}$), and places a special alternating hexadecimal marker byte (0x05 / 0xFA) in the upper 8 bits. A DoP-compatible DAC detects the 0x05/0xFA marker, strips it away, and routes the untouched 16 bits directly into its native 1-bit DSD hardware path—with zero conversion to PCM!

Ironically, the historic debate between “Multi-Bit PCM” and “1-Bit DSD” has largely converged inside modern silicon converter chips (such as ESS Sabre, AKM Velvet Sound, and Cirrus Logic MasterHIFI). Pure 1-bit DACs suffer from high sensitivity to clock jitter and non-linear inter-symbol interference (ISI), while pure 24-bit R-2R ladder DACs cannot maintain $0.00001\%$ resistor matching accuracy across 24 bits.

Consequently, virtually all modern flagship studio ADCs and DACs are Multi-Bit Delta-Sigma Converters: whether you feed them a 24-bit / $96\text{ kHz}$ PCM file or a 1-bit DSD256 file, their internal architecture converts the signal into a 5-bit or 6-bit ($32$ to $64$ level) Delta-Sigma modulator running at $11.2\text{ MHz}$ to $45\text{ MHz}$, paired with Dynamic Element Matching (DEM) that scrambles tiny resistor mismatches into benign white noise. Only boutique discrete 1-bit DACs (such as Chord Electronics’ FPGA Pulse Array, Playback Designs, or T+A’s dedicated True 1-Bit converter) maintain completely separate discrete 1-bit vs PCM conversion paths.