# producer

One-click narration and podcast mastering. Give it a raw recording; it denoises,
enhances, applies a warm "audiobook narrator" voice chain, and delivers a
loudness-normalized master.

```bash
./producer episode.wav
# -> episode_processed.wav (44.1 kHz mono, RMS -20 dB, true peak <= -3 dB)
```

Everything a user doesn't need to touch lives in `lib/`: the first run
bootstraps its own CPython 3.11 environment (via [uv](https://docs.astral.sh/uv/))
into `lib/.venv` and installs pinned dependencies. No system packages, no
manual venv, no configuration required.
`./producer doctor` checks the environment at any time.

## The sound

The default `audiobook` profile targets the classic close-narration master:
high-pass at 80 Hz, a gentle mud cut, low-shelf warmth, serial compression
(2:1 then 3:1), de-essing, restrained presence/air, breath ducking, then
normalization to RMS -20 dB with a -3 dB true-peak ceiling.

`--profile podcast` switches to broadcast loudness: -16 LUFS integrated,
-1.5 dBTP ceiling, 48 kHz, slightly brighter EQ. `--profile radio` targets the
deep, warm broadcast voice: +3 dB low shelf at 100 Hz, a tighter mud cut, heavy
serial compression, dynamic resonance control (`soothe`) that ducks boxy
200-450 Hz and harsh 2.5-6 kHz peaks only while they stick out, and asymmetric
tape saturation (`tape`) for the even-harmonic analog sheen.

## Engines

Every AI stage is swappable; the trade-offs are explicit:

| Stage | Choices | Notes |
|---|---|---|
| `--denoise` | `dfn3` (default), `zipenhancer`, `off` | [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet): 48 kHz full-band, faithful, fast on CPU, CUDA optional. [ZipEnhancer](https://github.com/gyj1201/zipEnhancer) (ICASSP 2025 SOTA, PESQ 3.69): 16 kHz native, bandwidth is restored after, slightly softer highs. |
| `--enhance` | `off` (default), `mossformer2`, `resemble` | [MossFormer2_SE_48K](https://github.com/modelscope/ClearerVoice-Studio) (ClearVoice): full-band studio restoration, GPU-strong. [Resemble Enhance](https://github.com/resemble-ai/resemble-enhance): generative restoration for badly damaged audio, runs in an isolated venv; can alter voice timbre, so it is opt-in. |

Engine dependencies (torch ~2.5 GB on CUDA, model weights ~8 MB) are installed
lazily on first use into `lib/` (`lib/models`, `lib/venvs`). Prefer
`--denoise off` for pure-DSP masters with zero heavy downloads.

## Usage

```bash
./producer in.wav [-o out.wav] [--profile audiobook|podcast|radio]
    [--denoise dfn3|zipenhancer|off] [--denoise-strength 0-1]
    [--enhance off|mossformer2|resemble] [--enhance-strength 0-1]
    [--no-dsp] [--no-levelling]
    [--hpf-hz N] [--mud N] [--warmth N] [--soothe N] [--compress N]
    [--tape N] [--deess N]
    [--presence N] [--air N] [--breath N]          # 0-1 strength each
    [--target N] [--ceiling N]                     # loudness target / TP ceiling (dB)
    [--format wav|flac|mp3] [--sample-rate N] [--bit-depth 16|24|32]
    [--device auto|cuda|cpu] [--batch] [--report] [--dry-run] [-v]
./producer doctor
```

Batch: `./producer --batch takes/ -o masters/` expands a directory (or glob)
and processes each file. Reports: `--report` writes `<output>.report.json`
with before/after RMS, true peak, LUFS, noise floor, and per-stage timings.

Without `-o`, outputs are written next to the input as
`<input>_processed.<format>`. If the output path already exists, producer
prompts to overwrite, rename (auto-numbered `..._1`, `..._2`, ...), or cancel;
non-interactive runs (no terminal on stdin) auto-rename and say so.

Persistent settings go in `config.toml` in the repository root (auto-created on
first run; CLI flags always win). Example:

```toml
profile = "audiobook"

[denoise]
engine = "dfn3"
strength = 1.0

[audiobook]
air = 0.4
deess = 0.8
```

## Requirements

- Linux (glibc: Arch, Debian, Fedora...), macOS, or Windows/WSL
- `curl` (bootstraps uv; removed afterwards is fine)
- `ffmpeg` on PATH for MP3 output and non-soundfile inputs (WAV/FLAC/OGG
  work without it)
- NVIDIA GPU optional; everything falls back to CPU

## Development

```bash
./producer doctor                      # environment check
lib/.venv/bin/python -m pytest lib/tests
lib/.venv/bin/python -m pytest lib/tests -m slow   # engine integration (needs deps)
lib/.venv/bin/ruff check lib/src lib/tests
```

Engine wrappers are exercised against real models by the `slow` tests
(`PRODUCER_TEST_RESEMBLE=1` additionally opts into the isolated generative
venv). Unit tests for meters, DSP, loudness, pipeline, CLI, and I/O run
dependency-free.

## License

MIT — see [LICENSE.md](LICENSE.md). Model weights keep their upstream licenses
(DeepFilterNet MIT/Apache-2.0, ClearVoice Apache-2.0, Resemble Enhance MIT,
ZipEnhancer MIT).