Follow
Audiophile Inventory on Facebook Audiophile Inventory on Google+ Audiophile Inventory on Twitter Audiophile Inventory on Instagramm Audiophile Inventory on Pinterest Audiophile Inventory on Telegram Audiophile Inventory on YouTube

DSD DSF DFF Editor Software [Audio Quality Issues, How It Work]

Share 

Audio Basis - educational articles

DSD editing software is intended to altering DSF and DFF audio file content: gain, cutting, merging, EQ, etc. DSD is 1-bit signal based on sigma-delta modulation and its editing is not obvious at first glance unlike PCM. There are many disputes about "native" and "non-native" editing, damage of original stuff. Read how to edit DSD and about real damage of the editing.

 

DSD and PCM comparison

 

DSD and PCM oscillogramms

DSD and PCM oscillogramms

 

It is necessary to remember, that digital signal isn't "stairs", because the "stairs" is filtered by low frequency filter. And analog signal, restored from digital one is smooth.

We can imagine how PCM editor work. It apply math on each sample or group of samples of musical signal.

As example, to boost level we can multiply each sample of PCM to gain value.

DSD is 1-bit signal and it is difficult to understand what is possibly to do with 1-bit sequence.

When DSD sample is multiplied to gain value, the result value is different from both 0 and 1. And it isn't possible to store the value as 0 or 1.

To easier introducing into DSD editor matters, it is need to find out total base for one- and multi-bit stuff.

In general, DSD and PCM are the same. But there are several differences:

DSD PCM
Only part of the frequency band [sample rate]/2 is used for audio Full frequency band [sample rate]/2 is used for audio
Noise shaping is mandatory to achieve appropriate signal/noise ratio in useful audio band Noise shaping isn't used as rule
Noise energy is pushed to high frequency range and have value comparable with useful signal Noise is distributed evenly throughout full frequency band (0 ... [sample rate]/2)
Stabilty issues of sigma-delta modulator (noise shaper) due its input overload No stability issues
Audible products of intermodulation are more probable due more significant ultrasound noise level Lesser probability of audible intermodulation distortions. However, significant ultrasound harmonoics may be dangerous

 

At the picture, spectrums of a quantization noise are depicted.

 

PCM and DSD transformation. What is difference

PCM and DSD. What is difference

 

At left part of the picture PCM quantization noise is depicted. At the middle picture bit depth is decreased. At right part of the picture noise shaping applied to the quantization noise. This technique is used for DSD to improve signal-noise ratio in the audible band.

Read more details about DSD vs PCM here >

Comparison DSD vs FLAC (infographic) read here >

 

 

DSD editing is like to PCM?

 

DSD have bit depth 1 bit.

But:

  1. Modern computers "natively" perform multi-bit calculations only.
  2. Adding of two 1-bit numbers demand 2 bits for result.
  3. What about result bit depth of non-integer multiplication (dB scaling as example)?

Therefore, for editing of 1-bit DSD sample may be converted to multi-bit sample for compatibility with processor unit. Further different processing may be applied like PCM.

Edited stuff must be reduced back to 1-bit. It may be implemented as sigma-delta re-modulation.

 

 

DSD editor inside

 

Simple scheme of DSD editor.
Upper part of the picture: spectrums per each stage are shown

 

Simple DSD editor

 

In this scheme high-frequency "legacy" DSD noise come to editing. It can cause:

  • overload, that lead to broken stability of output sigma delta modulator (re-modulator),
  • audible products of intermodulation distortions (audible noise).

DXD PCM format is partitially same such 1-bit to multi-bit conversion. But, there is filtering with decimation is performed. And DXD have integer format of samples.

 

Decimation is multiple sample rate division by deleting samples. Filtering should be applied before decimation. Because high frequencies may become audible if decimated signal.

 

Sample format is matter of target signal/noise ratio. Any further processing should be "transparent" for target resolution. I'd recommend use 3...10 times better precision, than final bit depth. Such figures are same to measurement recommendations. But it is not hard rule.

For DSD, target precision is defined by noise level, that depend on quality of sigma-delta modulator and target sample rate. Therefore, to choose internal bit depth of the editor software, need to refer to available re-modulator's noise level at target sample rate.

Read details below.

 

DSD editor with filtering of "legacy" DSD noise at high frequencies

DSD editing with filtering

 

For such way, probability of overload and level of audible intermodulation distortion products are decreased. Because lesser ultrasound content generate lesser intermodulation audible products.

 

DSD editor with filtering of "legacy" DSD noise and decimation

DSD editing with filtering and decimation

 

Decimation allow to save computing-resource consuming (time of processing). Because sample number, that used for calculations, is lesser. However, simple processing may have lesser computing profit, than resources, spent to decimation and upsampling.

As example, if you perform gain altering only, this operation don't consume many time. Upsampling time losses may be more, than profit by decimation. But filtering give advantages, described in the previous case. And decimation reduce filtering computation number.

Therefore, during desing need to acount these things.

 

 

What issues is caused by processing of multi-bit DSD audio?

 

When 1-bit audio sample is converted to multi-bit one, it keep noise energy comparable with useful audio signal at high frequencies. It is excess energy. It does not contains information.

But the energy can cause:

  • overload;
  • intermodulation distortions.

These 2 things are most dangerous for sound quality in DSD editing.

Overload can cause stability disrupting of the re-modulator.

Intermodulation distortions can give audible products despite the DSD-noise is ultrasound. Especially need to account, that the noise level is comparabnle with useful audio signal.

 

 

DSD processing types

 

Mixer (sum)

Mixing is a sum of original 1-bit samples, transformed to multi-bit values. Сonversion to 1-bit value is applied via sigma-delta re-modulation after summation.

DSD mixing audio

DSD mixing audio

 

As example, 2 signals are mixed. Such operation sum full frequency bands with width [sample rate]/2 of both original signals.

Maximal level may growth up to 2 times. Also noise energy is growth up to 2 times.

"Up to" because sum level depend on phase of the original signals. If 2 identical signals are summarized with zero time shifting, total level 2 times more than each of input signals.

Noise is stochastic process. So two noise sum give stochastic value too. Double level value also probable.

If original signal have 0 dB (maximal) level, summation the signal with any other one cause input overload of the sigma-delta re-modulator.

Even if no overload then summing useful audio signals, summation of the high-frequency noise of both signals can cause overload.

To avoid the overload need use filtering.

 

Gain altering (multiplication)

Gain altering is multiplication. Multiplication may be either integer or float point. Gain altering step of audio applications in range 0.1 ... 2 dB, as rule.

DSD gain altering

DSD gain altering

 

To keep output signal precision, we can try to use only integer calculations with expanding bit resolution. But float point formats are significantly easier for programming. However, integer formats may be preferable for processing implementation on field-programmable gate array (FPGA) chips.

Overload isn't defined by level of separate spectral component, like music. But overload is defined by total energy of the spectrum, that include not only useful audio signal, but "legacy" DSD noise.

Even if useful audio signal is lesser than maximal level, the total energy can cause overload of the re-modulator. Because there is the "legacy" noise at high frequencies.

Gain reducing can some decrease probability of overload, because the "legacy" noise is reduced too. But level boosting give higher overload probability.

When the noise is filtered, overload is solely defined by useful audio signal.

There is opinion, that digital filtering is destructive, because it generate ringing artefacts after filter. But it is balance between the artifacts and maximal available level after DSD editing.

However, author don't know about researches of impact ringing to perceived audio quality. And ringing is matter of steep level transients rather. But musical signal is smooth, as rule.

If we use filter with "deep" suppression in stop-band, we can apply decimation of samples to decrease sample rate and faster processing. Smooth filter can cause additional audible products transferred from high frequencies of original spectrum.

From point of view of my development experience, these matters is reference points only. And we can look for hidden opportunities around these points to achieve our aims.

 

Overdrive, tube and tape emulators, other enhancers, compressor (non-linear processing)

These kinds of processing are pure non-linear processing. When signal with high frequency noise pass thru non-linear circuit, the noise can generate audible products due intermodulation distortions. Main trouble here is noise energy that comparable with useful signal. Thus the audible intermodulation products may have significant level.

DSD enhancers

DSD enhancers

 

Therefore, non-linear processing can cause audible noise. Similar example we can watch in the video.

Compressing can not be considered as pure non-linear processing. Because in defined level range compressor is linear for momentary level value. However, linearity is changed during time of compressor gain switching. When the gain is switched and constant at new value, compressor work as linear device.

 

Video: Removing of audible products of intermodulation distortions

 

For correct non-linear processing, "legacy" DSD noise should be suppressed before the processing. Otherwise audible intermodulation products of high frequency noise may have significant value.

 

Equalizer (amplitude-frequency correction of spectrum)

Equalizer (EQ) is linear processing. However, it can cause overload when equalization band is boosted. The overload can cause broken stability.

DSD equalizer audio

DSD equalizer audio

 

Reverberation (sum)

Reverberation is linear processing in general case. But there summation and multiplication are used. Also EQ may be applied. Convolution algorithm may have different effect to level and spectral energy. So overload at input of sigma-delta remodulator may happen.

DSD reverberator

DSD reverberator

 

Merging and spliting

These operations are lossless. Because it is only manipulations under samples without changing.

DSD splitting

DSD splitting

 

DSD merging

DSD merging

 

 

How to edit DSD

 

There are 2 ways:

  1. Using of integrated editor (1-bit/multi-bit converter [optionally: filtering and decimation] + processing + [optionally: upsampling] + re-modulator).
  2. Using conversion DSD to PCM > editing PCM in audio editor > conversion PCM back to DSD.

 

DSD editing ways

DSD editing ways

 

When converter is used, it is recommended to use PCM formats with 32- or 64-bit integer or float bit depth to "transparent" (for music) processing. Float point format have better overload abilities comparing integer ones.

Decimation in the converter can improve processing performance and compatibility with PCM editor software.

Steeper filter cause more ringing artefacts. But decimation after sloping filter can cause additional audible products by high frequency components.

Sample rate must provide 0 ... 20 kHz useful band.

Depending on converter implementation, higher sample rates can cause additional "legacy" DSD noise due wider band. But some conversion software allow to choose passed band of input 1-bit signal, to reduce impact the modulation noise.

For PCM sample frequencies multiple 44.1 kHz, time of conversion (DSD to PCM) may be lesser, than for sample frequencies based on 48 kHz.

General recommendation to choose the conversion software is comparing by level difference between overload and noise/aliases (maximal dynamic range).

Read more about multiple vs non-multiple resampling here >

 

 

Conclusions

 

  1. Before processing 1-bit samples should be converted to multi-bit ones.
  2. After processing multi-bit samples need to re-modulate back to 1-bit ones.
  3. Overload after processing is dangerous for the re-modulator.
  4. Non-filtered DSD "legacy" noise (after conversion 1-bit to multibit samples) can cause input overload of the re-modulator.
  5. For non-linear processing the non-filtered noise can cause audible products (intermodulation distortions).
  6. Losses of DSD editing without processing is similar to PCM-resampling.
  7. Decimation can save time of processing, but cause additional noise for sloping filters before decimation.
  8. DSD editing may be performed via either integrated DSD editor or DSD to PCM conversion, usual PCM editor and PCM to DSD back conversion.

 

References:

  1. Josh Reiss, Mark Sandler. Digital audio effects applied directly on a DSD bitstream: link

 


Read the articles
Share 

Copyright © Yuri Korzunov
["Audiophile Inventory" (since 2011), "AudiVentory", "Audiophile Inventory by AudiVentory.com".
Также "Audiophile Inventory by AudiVentory.com" (прежде "Audiophile Inventory") охраняемое коммерческое обозначение],
2010-2018. All Rights Reserved.

Site map      Terms and Conditions