AudioCraft Audio Quality: Clean Chunk Clicks and EnCodec Hiss

A long Meta AudioCraft render can sound convincing until one tiny tick repeats at a boundary, a pale hiss follows the cymbals, or the chorus develops a grainy top layer. Those are useful listening clues. They are not proof that AudioCraft itself created the defect. Your checkpoint, decoder, extension settings, wrapper, edits, and later encoding all affect the file that reaches your speakers.

The short practical answer is to preserve the first WAV, mark the exact timestamps, and separate local discontinuities from texture that moves through the full mix. Repair one real click with a short crossfade. Treat repeating hiss or grit only after you can hear it in both an exposed passage and a dense one. Do that cleanup before mastering, because a limiter can make quiet roughness more obvious.

Checked September 21, 2026. The current official AudioCraft repository describes a research framework that includes MusicGen, AudioGen, and EnCodec, alongside other models and decoders. Its MusicGen documentation uses a 32 kHz EnCodec tokenizer with four codebooks sampled at 50 Hz. The official demo saves generated audio as WAV and does not document native MP3, FLAC, or separated-stem output as a standard MusicGen export.

The short answer: optimizing AudioCraft audio fidelity

Start with the rendered file, not a generic preset. If the problem is one sharp tick at an edit, fix that event locally. If a thin metallic layer rides across vocals, cymbals, and reverb, compare a restrained cleanup pass with the original at matched loudness. If the melody, rhythm, instrument choice, or structure is wrong, return to generation or editing instead of processing the whole mix.

MusicGen can create up to 30 seconds in one pass. For longer durations, the official implementation carries earlier token context forward. Its default extension settings use an 18-second stride with 12 seconds of overlap in the demo description. That is not the same as blindly joining isolated five-second WAV blocks, and it does not guarantee audible seams.

I listen to a long render once without watching the timeline. Then I mark the moments that pull my attention away from the music. That order prevents a grid line or spectrogram color from turning into a problem I never heard.

Use the AI music artifacts guide when you need to separate a steady noise floor from moving shimmer, wobble, or smeared transients. Each symptom calls for a different test. Calling every bright defect “codec hiss” may sound precise, but it can send you toward the wrong tool.

Diagnosing AudioCraft audible defects

Loop two or three seconds around the first problem, then compare it with a clean section. Write one plain-language note before changing anything: “single tick at 36.0 seconds,” “sand follows the hi-hat decay,” or “bass weakens in mono.” A repeatable description is more useful than a guessed internal cause.

• Boundary tick or click: A very short impulse appears at a repeatable timestamp. Zoom into the waveform. Look for a sudden level jump, a cut that misses a zero crossing, or a left/right channel discontinuity. Check whether the interval matches your own extension or slicing settings.
• High-frequency grit and haze: Cymbals, reverb, or bright synths develop a sandy layer that moves with the musical source. This can resemble codec reconstruction noise, but the sound alone does not identify the exact decoder or codebook.
• Steady hiss: A similar “shhh” remains in gaps and tails. Compare the intro, a pause, and the final decay. If the texture changes with the arrangement, it may be moving artifact energy rather than ordinary stationary hiss.
• Midrange resonance: A vocal, synth, or acoustic layer becomes sharp only on certain notes. A narrow dynamic EQ test may help, but a broad cut can thin every source around the defect.
• Low-end instability: Sub energy swells, vanishes in mono, or triggers a limiter without adding useful weight. Compare headphones, speakers, and mono before filtering.
• Transient smearing: A snare or pluck keeps its body but loses a clear front edge. Do not confuse a deliberately soft attack with damage. Compare several hits rather than judging one event.
• Musical failure: A wrong note, broken lyric, abrupt ending, or incoherent section is not an audio-cleanup problem. Edit, extend again, or regenerate it.

A spectrogram can show a click as a brief vertical line and steady hiss as a broad horizontal cloud. It cannot tell you whether that energy is unwanted or which model stage produced it. The waveform, spectrogram, and your ears should point to the same timestamp before you act.

For upper-mid pain, the harsh-highs guide explains a safe listening test and a clear stop condition. If the whole mix becomes darker before the harsh moment settles, undo the move.

Source preflight: export formats and stem isolation

Record the exact path that produced the audio: AudioCraft commit or package version, model checkpoint, generation duration, extension stride, decoder, seed when available, wrapper, and every conversion after generation. Two interfaces can call the same model and still save different channel layouts, levels, or file formats.

The official MusicGen documentation describes a 32 kHz EnCodec tokenizer with four codebooks sampled at 50 Hz. In plain language, the waveform is represented as several streams of discrete audio tokens before it is decoded back to sound. Fewer bits or a different decoder can change reconstruction texture, but hearing hiss does not prove that one codebook is at fault.

The repository demo exposes WAV output. It also offers the default GAN-based compression decoder and an optional MultiBand Diffusion decoder, noting that the default path can show artifacts on sounds such as crashes and snares while the diffusion path costs more compute. That is a documented comparison, not a promise that diffusion will repair every file.

AudioCraft is a framework rather than one hosted download product. The official materials do not promise a native MP3 button, FLAC export, a universal 24-bit setting, or separated stems for standard MusicGen generation. A third-party interface may add those functions. Treat them as features of that interface and verify whether a “stem” is truly generated separately or separated later from a stereo mix.

Preserve the original generated WAV before normalization, conversion, or mastering. If your wrapper only gives you MP3, keep that untouched file and avoid converting it repeatedly. Turning MP3 into WAV changes the container, not the detail already removed by lossy encoding.

Check rights with the same precision. The repository code is released under MIT. The published MusicGen model weights are CC-BY-NC 4.0. Those are different layers. The exact checkpoint, conditioning audio, samples, wrapper terms, and intended use still matter.

AudioCraft chunk overlap clicks at 5-second intervals

AudioCraft Chunk Overlap Clicks at 5-Second Intervals

The five-second grid below is a diagnostic example for a file that appears to click on that schedule. AudioCraft does not use fixed five-second chunks by default. The official MusicGen API supports up to 30 seconds in one pass, while the demo describes an 18-second stride and 12 seconds of overlap for longer generation.

Timestamp	What to inspect	Evidence of a real boundary problem	Next action
0:05	Waveform and both channels around the marker	A sharp level or DC jump at exactly 5.000 seconds	Check custom wrapper chunk or slicing settings
0:10	Same zoom and listening level	A matching impulse with the same polarity or shape	Bypass later batch processing and render again
0:15	Spectrogram plus waveform	Brief broadband vertical line that is also audible	Repair locally with a very short crossfade
0:20	Left/right channels separately	Click occurs in one channel or phase flips across both	Correct the channel edit; recheck mono
0:25	Unprocessed render versus converted file	Original is clean but converted file clicks	Fix the conversion or concatenation stage
0:30	Official one-pass boundary for common checkpoints	Click exists only after an extension or edit	Compare extension settings and preserve more context

A grid can reveal regularity, but it can also create confirmation bias. Nudge the marker away from the expected timestamp and listen blind. If the “click” disappears when you stop looking at the grid, do not process it.

The official long-form code concatenates generated token sections while retaining context from the preceding section. It is not documented as a waveform crossfade at every five seconds. A custom batch script may still split, decode, normalize, or join audio in a way that creates a discontinuity. Test each stage rather than blaming the framework.

I keep one render from before batch conversion whenever possible. It gives me a clean comparison when a click appears later. Without that reference, it is easy to repair the final file while leaving the broken export step in place.

Restrained manual DAW path

Duplicate the untouched file. Put the original and working copy on aligned tracks, route them to the same output, and disable any automatic loudness normalization. Make one change at a time.

Repair an isolated click with a short crossfade. Zoom until you can see the waveform cycles. If there is an actual discontinuity, place a tiny equal-power crossfade around it or redraw the single sample transition with a suitable repair tool. Keep the fade shorter than the musical attack. A long fade across a snare can remove the click by removing the snare.

For sub-rumble, test a high-pass filter below the useful bass. A starting point around 20–30 Hz can reveal whether inaudible energy is stealing headroom, but it is not a universal cutoff. Raise it slowly while the full mix plays and back off when the kick or bass loses weight.

For a narrow resonance, use a dynamic EQ so the band lowers only when the harsh note appears. Begin with a small range and moderate timing. Bypass often. If the vocal air, cymbal detail, or room tone disappears with the resonance, the move is too broad or too deep.

For steady hiss, use the least reduction that makes the noise less distracting. Listen to the removed signal when your tool provides a residue mode. Recognizable vocals, cymbals, or reverb in the residue mean the process is taking music too.

For moving grain across several sources, static EQ may darken the file without separating artifact from music. Try a restrained full-mix cleanup candidate, then check the busiest section, an exposed tail, and the smallest transient. Stop when attacks soften, ambience closes, or stereo width shrinks.

Batch processing is appropriate only after one representative file passes this review. A script can repeat settings consistently; it cannot decide that a different song needs regeneration rather than restoration. Store the source path, settings, and output separately so you can reverse the operation.

Compare at matched loudness after every meaningful step. A louder version often feels clearer and more exciting even when its roughness is worse. The cleanup-versus-mastering guide explains why source repair should come before final loudness and delivery work.

The Sunofix cleanup path

I built Sunofix for the point where the song already works but a synthetic edge still follows several sources through the exported mix. A single boundary click is usually faster to repair locally. Sunofix becomes more relevant when hiss, grain, brittle highs, or moving haze continues across different passages and broad manual processing starts to remove wanted music.

Upload the best lawful WAV or MP3 you have, choose the dominant audible problem, and keep the untouched source nearby. Compare the cleaned WAV with the original at the same perceived loudness. Listen first to the named defect, then to a full verse and chorus. The result is useful only when the distraction recedes while the arrangement, timing, transients, stereo space, and emotional movement remain intact.

Clean an AudioCraft render in Sunofix when the unwanted texture moves through the full mix and a narrow local repair is no longer enough.

Frequency diagnostics and before-and-after playback help you verify the change. They are evidence, not a score. If a smoother spectrogram comes with dull cymbals or a smaller chorus, keep the original or choose a gentler pass.

Sunofix does not know the private history of your render and does not need to claim that AudioCraft caused the symptom. It processes the file you provide. It does not generate new stems, change model settings, replace lyrics, correct a wrong melody, rebuild an arrangement, or master the final release.

Clear technical and legal boundaries

“EnCodec codebook hiss” is a useful search phrase, but it should remain a listening hypothesis unless you reproduce the sound across controlled encode/decode settings. EnCodec uses residual vector quantization codebooks, and MusicGen predicts several token streams. That architecture does not make every bright texture a codebook defect.

The same caution applies to “chunk stitching clicks.” An audible discontinuity may come from extended generation, but it may also come from a custom wrapper, file concatenation, normalization per segment, a DAW edit, sample-rate conversion, or lossy encoding. One file does not prove that AudioCraft caused it.

Audio cleanup cannot restore samples already lost to clipping, codec reconstruction, or an earlier destructive edit. It cannot repair lyrics, melody, arrangement, or performance. It cannot create a clean isolated stem from a finished stereo mix or recover an original studio recording that never existed.

Mastering remains a separate stage for final tonal balance, dynamics, loudness, sequencing, and delivery quality control. Do not use a limiter to hide hiss or a seam click. Clean the source first, then master the cleaner candidate.

The MIT code license does not override the CC-BY-NC 4.0 model-weights license. Cleanup also does not grant legal clearance or guarantee distributor approval. Review the exact model card and license, the rights to prompts and conditioning audio, any third-party interface terms, and the rules of the destination where you intend to publish.

AudioCraft audio release checklist

1. Record the AudioCraft version, checkpoint, decoder, seed, duration, extension stride, wrapper, and later conversions.
2. Preserve the original generated WAV or the earliest untouched file available.
3. Mark each audible problem by timestamp before opening processors.
4. Check whether clicks repeat at your actual extension, slice, or batch interval.
5. Compare the one-pass render with the extended or concatenated file when possible.
6. Repair an isolated click with a short crossfade instead of processing the whole song.
7. Inspect both channels and compare stereo with mono at every suspected boundary.
8. Treat sub-rumble only when it steals headroom or destabilizes downstream processing.
9. Use dynamic EQ only when a repeatable resonance becomes harsh.
10. Reduce steady hiss gently and listen for wanted music in the residue.
11. Test broader cleanup only when unwanted texture travels through several sources.
12. Compare every candidate at matched loudness with the untouched source.
13. Stop when cymbals dull, attacks soften, ambience closes, or the mix narrows.
14. Regenerate or edit wrong notes, lyrics, timing, structure, and abrupt musical endings.
15. Master only after the cleanup candidate preserves the musical idea.
16. Recheck the exact code, weights, inputs, wrapper, and release terms before distribution.

A clean result is not the file with the flattest spectrum or the fewest visible lines. It is the version where the click, hiss, or grain stops stealing attention and the music still feels like the same piece. Keep the smallest useful change, leave enough headroom for mastering, and retain the original so every decision remains reversible.