What this chapter helps you decide
This chapter combines official technical material with a proposed production exercise. Generation, performance, pricing and listening comparisons have not been run. It is a supplement to the creative production course and can also be read independently.
- Compare preparation and constraints for local and hosted generation.
- Distinguish sample rate from codec.
- Inspect the audio exported for a video.
- 1Purpose and permissions
- 2Local or hosted constraints
- 3Approved export path
- 1Source format
- 2Convert a copy
- 3Inspect actual duration and synchronization
- 1Archive source
- 2Preserve unchanged
- 1Release copy
- 2Match delivery specification
This is an original conceptual workflow. The arrows show review order, not a product’s hidden architecture; a failure sends the work back to the responsible stage.
The execution location changes your responsibilities
A local workflow leaves the model, runtime, memory and storage under your control. A hosted service such as Suno supplies the execution environment, while available models, features and export conditions depend on the plan and date. Local does not automatically mean free or commercially unrestricted, and hosted does not automatically mean faster. Evaluate each candidate against the same purpose and record unknown performance as unknown.
MusicGen is one research example
The official MusicGen documentation describes text input and, for some variants, melody guidance, with a 32 kHz audio representation and distinct mono/stereo variants. Its guidance for medium models requires at least 16 GB of GPU memory; this is provider guidance, not a measurement on this computer or a Mac compatibility claim. The model card distinguishes MIT code from CC-BY-NC 4.0 weights. Inspect the applicable terms separately rather than treating the code licence as permission to commercially use the weights.
Complete the input/output table before execution
Record model and version, support for text, melody or lyrics, maximum duration, channel layout, output format and editing features. For Suno, check the features visible to the actual account and the approved download path. Listening in an interface does not establish that an export suitable for external use is available. The download FAQ, revised August 10, 2026, describes a September 3 change that includes older songs; check the current account allocation rather than copying a historical unlimited-download assumption.
Separate WAV, MP3, AAC and MP4
WAV is a container; when it holds PCM, it can serve as uncompressed editing audio. MP3 and AAC are audio coding formats; MP4 is a container that can hold video and audio. Renaming a filename does not convert its contents. Keep a lossless editing source when available, and avoid repeated lossy encoding. Sample rate is samples per second, bit depth is the precision of values, and channels are separate audio streams.
Resampling does not recover missing information
Converting 32 kHz audio to 48 kHz can satisfy a delivery specification, but does not recover absent high-frequency detail or create genuine stereo content. Interpreting samples at the wrong rate can change speed and pitch; actual resampling is a different operation. Ableton’s Audio Fact Sheet describes sample-rate conversion as non-neutral. Record the source rate, use a deliberate conversion, and check the resulting file instead of claiming a measured improvement.
Peak and loudness answer different questions
Peak describes an instantaneous level; loudness summarizes perceived level over time. Inspect integrated LUFS, true peak and listening impressions separately. EBU R128 is a broadcast recommendation; it does not establish a mandatory YouTube LUFS target. This course therefore leaves measurements null and prioritizes the actual delivery specification and freedom from audible distortion.
Keep an archive source and a release copy
YouTube’s encoding guide lists MP4, H.264, AAC-LC among audio options, and 48 kHz. MP4 + H.264 + AAC-LC is this course’s general delivery example, not a universal mandatory format or a quality guarantee; the page also notes Content Manager partner scope. Preserve the source and inspect the beginning, ending and synchronization after conversion. Even a video made from a still image needs this inspection.
Use and revise the production record
The input is a brief naming the use, duration and destination, plus materials whose usage conditions can be checked. The outputs are the production note, review sheet and the exercise deliverable. Keep the selected original separate from the editing and release copies.
- Narrow the candidates using the intended publication and material permissions.
- Check the model or plan’s inputs, outputs and export conditions.
- Specify the environment and storage before execution.
- Record acquired duration, format, sample rate and channels.
- Preserve the editing source and convert a copy to the delivery specification.
- Check distortion, synchronization, silence and the ending after conversion.
| Symptom | Likely cause to investigate | Revision |
|---|---|---|
| Playback speed or pitch changes | The sample-rate interpretation changed without converting the data. | Read the file at its correct original rate and resample only when needed. |
| A cue can be played but cannot be delivered | Download conditions were checked only after generation. | Check the approved export conditions before generation. |
Choose a method before running it
Compare local MusicGen and Suno by inputs/outputs, preparation, editing, usage terms and export conditions. Leave speed and cost unexecuted or unknown.
Deliverable: A method comparison note and delivery-file specification.
Completion criterion: The choice follows from the intended use and constraints, without an invented performance or cost ranking.
Execution status: not-run. This chapter contains no generated audio, timing log, listening result or completed publication.
Review checklist
- The weights licence was read separately from the code licence.
- Upsampling is not described as a measured quality improvement.
- Measured values and official specifications are distinct.
A comparison table with no invented benchmark
| Question | Local MusicGen | Hosted Suno |
|---|---|---|
| Who prepares execution? | User manages the model, runtime and storage | Service manages generation; account conditions still need checking |
| What can be edited? | Depends on model output and the editing project you retain | Depends on the account’s actual features and exported files |
| What constrains release? | Inspect weights, code and material permissions separately | Inspect applicable terms, download path, plan and material provenance |
| Measured speed and cost | null / not-run | null / not-run |
An illustrative delivery spec might name a lossless editing source, a 48 kHz release copy when needed by the chosen video workflow, and an MP4/H.264/AAC-LC review file. Source and delivery rates are separate fields. The spec is a plan: actual channels, duration, loudness and conversion defects remain unmeasured.
MENTAL MODEL / SHOT DESIGN
Give each shot one job.
Total: 12 seconds. Define a reference image, a start state, one action, and an end state for each shot. Arrange the still images in this order before generation to find transitions that do not communicate your idea.
Sources
Publication dates belong to the source; access dates record when it was checked. Community observations are separate from official statements.