LW IT Solutions
17 articles in this category
Follow this category by RSSThe “Music Production” category covers the technical side of making music with a computer: digital audio workstations, virtual instruments, recording chains, mixing, mastering and the delivery of finished tracks to streaming services. Music production has long since become a software discipline in which plugin formats, CPU load, latency, sample rates and file formats matter as much as melody and arrangement. This section examines that toolchain with the same measuring approach used across this blog: versions, numbers and verifiable behaviour instead of marketing claims.
Articles in this category compare DAWs such as FL Studio, Ableton Live, Logic Pro, Cubase and Bitwig Studio, explain audio fundamentals like loudness normalisation (LUFS), true peak, buffer sizes and sample rates, and follow the fast-moving field of AI in music – from stem separation and AI session players inside the DAW to fully generated songs, their licensing situation and the rules streaming services apply to them. Practical tutorials deal with sound design, home studio setups, audio interfaces, MIDI and the metadata a release needs, such as ISRC codes and AI disclosure in the credits.
The goal is a solid basis for decisions for producers, beatmakers, sound designers and technically minded musicians: which software suits which workflow, where a studio setup loses quality or time, and how human-made and machine-generated music differ in control, rights and effort. Where a calculation helps – latency from buffer size, delay times from tempo, file sizes from sample rate and bit depth – it is shown step by step, so that every figure can be checked.
Lossy encoders add samples before and after the music. Measured with ffmpeg: an MP3 without its Info frame starts 1105 samples late, and two halves joined back together leave 37.57 ms of silence at the seam. With the fields honoured, MP3, Opus and Vorbis join seamlessly, while AAC still left a gap of 20.16 ms until the track duration was honoured as well.
Comparing a mix against a reference only works at equal loudness, measured on a synthetic 2:40 piece in three versions: matching by sample peak left 10.78 dB of loudness difference, matching on the chorus alone was 2.5 dB off, and a reference running through the own master chain gained 1.6 LU. After loudness matching, the dense master still differed from the mix by up to 3.1 dB per third-octave band.
A vocal chain in FL Studio step by step, with the effect of each step measured on a synthetic voice signal: a high-pass at 100 Hz removed 29.4 dB below 100 Hz and cost 0.79 dB at the fundamental, the same high-pass in front of the compressor kept the voice 1.34 dB louder and removed a 1.45 dB dip on every plosive, and compression widened the distance between sibilant and vowel by 2.92 to 4.43 dB.
Sample rate conversion from 48 to 44.1 kHz, measured with ffmpeg: the default resampler, also used by -ar 44100, lowered 20 kHz by 1.3 dB and let a tone at 22.5 kHz through as an alias at 21.6 kHz, attenuated by only 14 dB. soxr stayed within 0.01 dB up to 20 kHz and attenuated the alias by 137 dB. A master hard-clipped at −1 dBFS had a sample peak of up to +0.27 dBFS after conversion.
Impulse responses with decay times from 0.3 to 4.8 s were convolved with ffmpeg and evaluated with C50, C80 and modulation depth. The reverb level decides more than the length, a pre-delay of 80 ms lowers C50 from 12.60 to 10.00 dB instead of raising it, and a 200 Hz high-pass in the send cuts the low end left after the loop from −23.3 to −35.6 dB.
Dither when exporting to 16-bit, measured with ffmpeg: without dither, a sine at −96 dBFS lost 3.7 dB and its third harmonic was only 0.9 dB below the fundamental. RPDF dither removed the distortion but let the spread vary with the input between 0 and 0.5 LSB. TPDF kept both constant at the cost of 0.5 LSB of noise (−96.4 dBFS). Noise shaping lowered the noise below 4 kHz by up to 17 dB and raised the total noise by 12 to 24 dB.
Both denoisers were run over a clean programme in which there was nothing to remove, so that everything missing at the output is damage. The noise reduction control barely changes anything, while the noise floor control moves the result by 20.9 dB, and the best setting lands within 1.6 dB of the real noise floor every time. Below a noise level of 40 dB under the signal, the best achievable improvement is 0.15 dB.
Clipper or limiter for a loud master: a measurement with ffmpeg compares six methods with the same ceiling of −1 dBFS. A hard clipper without oversampling produced aliasing 25 dB below the fundamental on a sine and, at −8 LUFS, a true peak of +0.41 dBTP in a 16-times cross-check, although no sample exceeded −1 dBFS. After AAC at 256 kbit/s, clipped masters rose to +1.95 to +3.07 dBTP, the oversampled limiter only to +0.36 dBTP.
All 23 crossfade curves of ffmpeg measured: seven hold the level when both sides carry the same take, two hold it when they carry different recordings, and none holds both, so the wrong choice costs 3.01 dB either way. A cut exactly on a zero crossing still clicks at −37.44 dBFS when the slope reverses there, and the fade needed to bury the click does not depend on the frequency of the material.
Three ways of widening a mix, measured with ffmpeg on a 32 second piece: raising the side costs nothing in mono beyond its own level, 0.30 to 1.10 dB spread evenly across the spectrum, but does nothing at all to a mono source. Delay based widening reaches a correlation of 0.04 or tears eight twelfth-octave bands more than 6 dB out of the mono sum, the deepest 9.37 dB at 50 Hz.
English
Deutsch
Polski
