Add cli

[mhrice]

No description provided.

[Copilot]

Pull request overview

This PR adds a stable-audio command-line interface entrypoint for running Stable Audio 3 generation (text-to-audio, audio-to-audio, inpainting/continuation) without writing Python, along with unit tests and documentation.

Changes:

• Introduces stable_audio_3/cli.py implementing argparse-based generation and WAV output saving.
• Adds tests/test_cli.py covering CLI flag wiring via mocked model and torchaudio I/O.
• Updates README.md with CLI usage examples and adds the stable-audio console script in pyproject.toml.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
stable_audio_3/cli.py	New CLI implementation: argument parsing, model loading, generation invocation, and output saving.
tests/test_cli.py	New unit tests validating CLI flag wiring with mocked dependencies.
README.md	Documents CLI usage and adds a note about uv sync behavior.
pyproject.toml	Adds a stable-audio console script entrypoint.

Comments suppressed due to low confidence (2)

stable_audio_3/cli.py:225

• torchaudio.load() returns (waveform, sample_rate), but StableAudioModel.generate() expects audio inputs as (sample_rate, tensor) tuples (see StableAudioModel._encode_audio_input unpacking). Passing the raw torchaudio.load() result here will swap sample rate and waveform and break audio-to-audio/inpainting. Load as (waveform, sr) and then pass (sr, waveform) for both init_audio and inpaint_audio.

    init_audio = None
    if args.init_audio:
        init_audio = torchaudio.load(args.init_audio)

    inpaint_audio = None
    if args.inpaint_audio:
        inpaint_audio = torchaudio.load(args.inpaint_audio)

stable_audio_3/cli.py:195

• When --prompt/--duration/--negative-prompt are passed with multiple values, the CLI will pass lists into model.generate(). The model validates list lengths with assert len(value) == batch_size (in StableAudioModel._build_conditioning_dicts), which will raise an AssertionError (and can become a silent logic bug under python -O). Add explicit CLI-side validation (or auto-infer batch_size from the list lengths when the user didn’t set it) and fail with parser.error(...) on mismatches.

    # --- Build scalar / list args ---
    prompt = args.prompt[0] if len(args.prompt) == 1 else args.prompt
    negative_prompt = None
    if args.negative_prompt:
        negative_prompt = (
            args.negative_prompt[0]
            if len(args.negative_prompt) == 1
            else args.negative_prompt
        )
    duration = args.duration[0] if len(args.duration) == 1 else args.duration

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

Comments suppressed due to low confidence (1)

stable_audio_3/cli.py:199

• Same issue as prompts: when --duration is provided multiple times, the CLI passes a list through to StableAudioModel.generate, which asserts that list length == batch_size. Please validate the length against --batch-size (or require batch_size to match) and fail via parser.error so users get a clear CLI message instead of an AssertionError.

    duration = args.duration[0] if len(args.duration) == 1 else args.duration

[mhrice]

This is fine

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (2)

stable_audio_3/cli.py:188

• Inpaint region times are not validated for non-negativity or ordering. Negative seconds can produce negative sample indices (masking the end of the clip due to Python negative indexing), and end <= start becomes a silent no-op region. Validate each start/end pair after parsing (>=0 and end > start) and parser.error on invalid regions.

    # --- Validate inpaint args ---
    if (args.inpaint_starts is None) != (args.inpaint_ends is None):
        parser.error("--inpaint-start and --inpaint-end must both be provided together")
    if args.inpaint_starts and len(args.inpaint_starts) != len(args.inpaint_ends):
        parser.error(
            "--inpaint-start and --inpaint-end must be specified the same number of times"
        )
    if args.inpaint_starts and not args.inpaint_audio:
        parser.error("--inpaint-start/--inpaint-end require --inpaint-audio")
    if args.inpaint_audio and not args.inpaint_starts:
        parser.error("--inpaint-audio requires --inpaint-start and --inpaint-end")

stable_audio_3/cli.py:214

• --duration accepts zero/negative values. In StableAudioModel.generate, non-positive durations cause sample-size adaptation to fall back to the default sample_size, which can generate unexpectedly long audio instead of failing fast. Add a check that all provided durations are > 0 and raise parser.error otherwise.

    # --- Validate list-flag lengths against batch size ---
    if (
        args.negative_prompt
        and len(args.negative_prompt) > 1
        and len(args.negative_prompt) != batch_size
    ):
        parser.error(
            f"Got {len(args.negative_prompt)} --negative-prompt values but batch size is {batch_size}"
        )
    if len(args.duration) > 1 and len(args.duration) != batch_size:
        parser.error(
            f"Got {len(args.duration)} --duration values but batch size is {batch_size}"
        )