aboutsummaryrefslogtreecommitdiff
path: root/lib/project/src
diff options
context:
space:
mode:
Diffstat (limited to 'lib/project/src')
-rw-r--r--lib/project/src/voiceforge/__init__.py3
-rw-r--r--lib/project/src/voiceforge/__main__.py3
-rw-r--r--lib/project/src/voiceforge/ai.py157
-rw-r--r--lib/project/src/voiceforge/ai_worker.py161
-rw-r--r--lib/project/src/voiceforge/audio.py128
-rw-r--r--lib/project/src/voiceforge/cli.py193
-rw-r--r--lib/project/src/voiceforge/config.py118
-rw-r--r--lib/project/src/voiceforge/pipeline.py199
-rw-r--r--lib/project/src/voiceforge/progress.py40
-rw-r--r--lib/project/src/voiceforge/setup.py203
10 files changed, 1205 insertions, 0 deletions
diff --git a/lib/project/src/voiceforge/__init__.py b/lib/project/src/voiceforge/__init__.py
new file mode 100644
index 0000000..1d8663d
--- /dev/null
+++ b/lib/project/src/voiceforge/__init__.py
@@ -0,0 +1,3 @@
+"""VoiceForge: local speech cleanup and mastering."""
+
+__version__ = "0.2.0"
diff --git a/lib/project/src/voiceforge/__main__.py b/lib/project/src/voiceforge/__main__.py
new file mode 100644
index 0000000..4e28416
--- /dev/null
+++ b/lib/project/src/voiceforge/__main__.py
@@ -0,0 +1,3 @@
+from .cli import main
+
+main()
diff --git a/lib/project/src/voiceforge/ai.py b/lib/project/src/voiceforge/ai.py
new file mode 100644
index 0000000..493be65
--- /dev/null
+++ b/lib/project/src/voiceforge/ai.py
@@ -0,0 +1,157 @@
+"""Automatic, isolated DeepFilterNet3 inference (Linux, Python 3.11 worker).
+
+Public progress callback: (stage: str, completed: float | None, total: float | None).
+Known totals are audio frames or download bytes (identified by stage);
+unknown totals use elapsed seconds as completed.
+First use downloads Python, wheels, and upstream model weights into the
+VoiceForge data directory (VOICEFORGE_HOME prefix or the XDG data location).
+"""
+from __future__ import annotations
+
+import ctypes
+import fcntl
+import math
+import os
+from pathlib import Path
+import platform
+import shutil
+import time
+
+from .setup import Progress, data_dir, ensure_uv, run_process, tool_env
+
+
+def _cuda_available() -> bool:
+ # Probe the driver without importing Torch or installing CUDA in the host env.
+ if os.environ.get("CUDA_VISIBLE_DEVICES") in {"", "-1"}:
+ return False
+ try:
+ driver = ctypes.CDLL("libcuda.so.1")
+ count = ctypes.c_int()
+ return driver.cuInit(0) == 0 and driver.cuDeviceGetCount(ctypes.byref(count)) == 0 and count.value > 0
+ except (OSError, AttributeError):
+ return False
+
+
+def _flavor(device: str) -> str:
+ if device not in {"auto", "cpu", "cuda"}:
+ raise ValueError("device must be 'auto', 'cpu', or 'cuda'")
+ if device == "cuda" or (device == "auto" and _cuda_available()):
+ if platform.machine() not in {"x86_64", "AMD64"}:
+ raise RuntimeError("The pinned CUDA 12.1 backend requires Linux x86_64.")
+ return "cu121"
+ return "cpu"
+
+
+def _worker_env() -> dict[str, str]:
+ # Same isolation as tool_env(); kept as a named seam for worker launches.
+ return tool_env()
+
+
+def _interpreter_is_local(python: Path, root: Path) -> bool:
+ """Whether the environment's interpreter belongs to this prefix.
+
+ An installation duplicated by copying (while the original remained) keeps
+ interpreter links that resolve into the original directory. Such an
+ environment must be rebuilt, or the copy would silently depend on files
+ outside this prefix and break when the original is removed.
+ """
+ try:
+ return python.resolve().is_relative_to(root)
+ except OSError:
+ return False
+
+
+def ensure_ai(device: str = "auto", progress: Progress | None = None,
+ verify: bool = False) -> Path:
+ """Install/validate the backend and model; return its Python executable.
+
+ CPU and CUDA environments are separate. Explicit CUDA never falls back;
+ auto uses Torch's availability check in the selected worker. Downloads trust
+ PyPI, download.pytorch.org, Astral's Python distribution, and upstream DF.
+ verify forces a real self-test even when the environment is marked ready
+ ('setup' uses this); ordinary processing trusts the marker so a ready
+ environment does not pay for a second model load per run.
+ """
+ flavor = _flavor(device)
+ try:
+ libc = os.confstr("CS_GNU_LIBC_VERSION") or ""
+ except (ValueError, OSError):
+ libc = ""
+ if platform.system() != "Linux" or not libc.startswith("glibc ") or tuple(
+ int(part) for part in libc.split()[1].split(".")[:2]
+ ) < (2, 28):
+ raise RuntimeError("The pinned DeepFilterNet3 backend requires Linux with glibc >= 2.28 "
+ "(for example Ubuntu 22.04+ or Debian 12+). "
+ "Alpine/musl is not supported by its prebuilt wheels.")
+ root = data_dir()
+ root.mkdir(parents=True, exist_ok=True)
+ environment = root / f"ai-df-0.5.6-torch-2.5.1-{flavor}-py311-v1"
+ python = environment / "bin/python"
+ marker = environment / ".voiceforge-ready"
+ started = time.monotonic()
+ # Serialize bootstrap/model downloads, including calls from separate CLIs.
+ with (root / "ai.lock").open("a") as lock:
+ while True:
+ try:
+ fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
+ break
+ except BlockingIOError:
+ if progress:
+ progress("Waiting for AI setup", time.monotonic() - started, None)
+ time.sleep(0.25)
+ ready = (marker.is_file() and python.is_file()
+ and _interpreter_is_local(python, root))
+ if not ready:
+ # A moved, copied, or half-installed environment cannot be
+ # repaired in place: interpreter links reference absolute paths.
+ # Rebuild it; the shared wheel caches make this fast, usually
+ # offline.
+ if environment.exists():
+ shutil.rmtree(environment)
+ uv = ensure_uv(progress)
+ env = _worker_env()
+ run_process([uv, "venv", "--python", "3.11", "--managed-python",
+ "--relocatable", "--allow-existing", str(environment)],
+ "Preparing AI Python 3.11", progress, env=env)
+ run_process([uv, "pip", "install", "--python", str(python),
+ "--index-url", f"https://download.pytorch.org/whl/{flavor}",
+ "torch==2.5.1", "torchaudio==2.5.1"],
+ f"Installing PyTorch ({flavor})", progress, env=env)
+ run_process([uv, "pip", "install", "--python", str(python),
+ "--only-binary", ":all:",
+ "--index-url", "https://pypi.org/simple", "deepfilternet==0.5.6",
+ "numpy==1.26.4", "soundfile==0.12.1"],
+ "Installing DeepFilterNet3", progress, env=env)
+ # Check a freshly built environment (and any explicit setup) with a
+ # real model load and inference; this also downloads the weights.
+ if verify or not ready:
+ run_process([str(python), "-I", str(Path(__file__).with_name("ai_worker.py")),
+ "--device", device, "--model-dir", str(root / "models/df3-0.5.6"),
+ "--check"], "Checking DeepFilterNet3", progress,
+ env=_worker_env())
+ marker.touch()
+ return python
+
+
+def denoise(source: Path, target: Path, device: str = "auto", strength: float = 1.0,
+ progress: Progress | None = None) -> None:
+ """Denoise mono 48 kHz WAV to float WAV, preserving the exact sample count.
+
+ strength is a dry/wet mix in [0, 1]. The worker publishes output atomically;
+ errors/cancellation leave an existing target untouched. Chunked inference is
+ bounded-memory, not bit-identical to inference over an entire recording.
+ """
+ if not math.isfinite(strength) or not 0 <= strength <= 1:
+ raise ValueError("strength must be finite and between 0 and 1")
+ source, target = Path(source).resolve(), Path(target).resolve()
+ if source == target:
+ raise ValueError("source and target must be different files")
+ if not source.is_file():
+ raise FileNotFoundError(source)
+ if not target.parent.is_dir():
+ raise FileNotFoundError(target.parent)
+ python = ensure_ai(device, progress)
+ run_process([str(python), "-I", str(Path(__file__).with_name("ai_worker.py")),
+ "--device", device, "--source", str(source), "--target", str(target),
+ "--model-dir", str(data_dir() / "models/df3-0.5.6"),
+ "--strength", str(strength)], "Denoising", progress, env=_worker_env())
diff --git a/lib/project/src/voiceforge/ai_worker.py b/lib/project/src/voiceforge/ai_worker.py
new file mode 100644
index 0000000..99ec9f5
--- /dev/null
+++ b/lib/project/src/voiceforge/ai_worker.py
@@ -0,0 +1,161 @@
+"""Private script executed only by the isolated backend Python."""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+from pathlib import Path
+import signal
+import tempfile
+import urllib.request
+import zipfile
+
+
+def report(stage: str, completed: float | None = None, total: float | None = None) -> None:
+ print("VOICEFORGE_PROGRESS " + json.dumps([stage, completed, total]), flush=True)
+
+
+def ensure_model(directory: Path) -> Path:
+ """Install a release-pinned, hash-checked archive atomically under setup lock."""
+ model = directory / "DeepFilterNet3"
+ if (model / "config.ini").is_file() and (model / "checkpoints/model_120.ckpt.best").is_file():
+ return model
+ if directory.exists():
+ raise RuntimeError(f"Incomplete model directory: {directory}. Remove it and retry AI setup.")
+ directory.parent.mkdir(parents=True, exist_ok=True)
+ url = "https://raw.githubusercontent.com/Rikorose/DeepFilterNet/v0.5.6/models/DeepFilterNet3.zip"
+ expected = "49c52edc8947ae1f9bf50d81530beaf3a2c3245aeaf34b6f31ff535cd22284d2"
+ # Release archive size is fixed alongside its hash, even without Content-Length.
+ expected_bytes = 7986207
+ downloaded = 0
+ report("Downloading DeepFilterNet3 model (bytes)", 0, expected_bytes)
+ with tempfile.TemporaryDirectory(prefix=".df3-", dir=directory.parent) as staging:
+ archive = Path(staging) / "model.zip"
+ digest = hashlib.sha256()
+ with urllib.request.urlopen(url, timeout=60) as response, archive.open("wb") as output:
+ while chunk := response.read(256 * 1024):
+ downloaded += len(chunk)
+ if downloaded > expected_bytes:
+ raise RuntimeError("DeepFilterNet3 model download exceeds the expected size")
+ digest.update(chunk)
+ output.write(chunk)
+ report("Downloading DeepFilterNet3 model (bytes)", downloaded, expected_bytes)
+ if downloaded != expected_bytes or digest.hexdigest() != expected:
+ raise RuntimeError("DeepFilterNet3 model checksum mismatch; refusing to load weights")
+ report("Verifying and extracting DeepFilterNet3 model")
+ unpacked = Path(staging) / "unpacked"
+ with zipfile.ZipFile(archive) as bundle:
+ for member in bundle.infolist():
+ destination = (unpacked / member.filename).resolve()
+ if not destination.is_relative_to(unpacked.resolve()):
+ raise RuntimeError("Unsafe path in model archive")
+ bundle.extractall(unpacked)
+ os.replace(unpacked, directory)
+ return model
+
+
+def main() -> None:
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--device", choices=("auto", "cpu", "cuda"), required=True)
+ parser.add_argument("--check", action="store_true")
+ parser.add_argument("--model-dir", type=Path, required=True)
+ parser.add_argument("--source", type=Path)
+ parser.add_argument("--target", type=Path)
+ parser.add_argument("--strength", type=float, default=1.0)
+ args = parser.parse_args()
+
+ # SIGTERM from the parent takes the same cleanup path as Ctrl-C.
+ def terminate(signum, frame):
+ raise KeyboardInterrupt
+
+ signal.signal(signal.SIGTERM, terminate)
+
+ import numpy as np
+ import soundfile as sf
+ import torch
+
+ torch.set_num_threads(min(4, os.cpu_count() or 1))
+ available = torch.cuda.is_available()
+ if args.device == "cuda" and not available:
+ raise RuntimeError("CUDA was explicitly requested, but PyTorch cannot use it. "
+ "Check the NVIDIA driver and CUDA_VISIBLE_DEVICES, or select cpu.")
+ device = "cuda" if args.device != "cpu" and available else "cpu"
+ # df.utils.get_device() consults DEVICE on every call, including model creation.
+ os.environ["DEVICE"] = device
+ from df.enhance import enhance, init_df
+ from df.model import ModelParams
+ from libdf import DF
+
+ model_path = ensure_model(args.model_dir)
+ report(f"Loading DeepFilterNet3 ({device})")
+ model, state, _ = init_df(str(model_path), log_level="ERROR", log_file=None)
+ if state.sr() != 48000:
+ raise RuntimeError("DeepFilterNet3 model must use 48 kHz")
+ if args.check:
+ # Exercise both model execution and the compiled libdf/NumPy ABI.
+ sample = enhance(model, state, torch.zeros(1, 4800), pad=True)
+ if sample.shape != (1, 4800) or not torch.isfinite(sample).all():
+ raise RuntimeError("DeepFilterNet3 self-test failed")
+ report(f"DeepFilterNet3 ready ({device})", 1, 1)
+ return
+ if args.source is None or args.target is None:
+ parser.error("--source and --target are required unless --check is used")
+ if not np.isfinite(args.strength) or not 0 <= args.strength <= 1:
+ parser.error("--strength must be in [0, 1]")
+
+ temporary = None
+ try:
+ with sf.SoundFile(args.source) as source:
+ if source.samplerate != 48000 or source.channels != 1 or source.format not in {"WAV", "WAVEX", "RF64"}:
+ raise ValueError("AI input must be a mono 48 kHz WAV; convert it before denoising")
+ frames = len(source)
+ fd, name = tempfile.mkstemp(prefix=".voiceforge-ai-", suffix=".wav", dir=args.target.parent)
+ os.close(fd)
+ temporary = Path(name)
+ # RF64 avoids RIFF's 4 GiB limit on very long recordings.
+ output_format = "RF64" if frames * 4 > 0xFFFFFFFF - 4096 else "WAV"
+ with sf.SoundFile(temporary, "w", samplerate=48000, channels=1,
+ format=output_format, subtype="FLOAT") as target:
+ block, context, overlap = 480000, 96000, 2400
+ previous = None
+ params = ModelParams()
+ report(f"Denoising ({device})", 0, frames)
+ for start in range(0, frames, block):
+ end = min(start + block, frames)
+ left, right = max(0, start - context), min(frames, end + overlap + context)
+ source.seek(left)
+ audio = source.read(right - left, dtype="float32")
+ if not np.isfinite(audio).all():
+ raise ValueError("Input contains non-finite audio samples")
+ if args.strength:
+ # Fresh STFT state per window; context warms normalization and
+ # recurrent layers. Pad to a whole hop before delay compensation.
+ state = DF(sr=params.sr, fft_size=params.fft_size,
+ hop_size=params.hop_size, nb_bands=params.nb_erb,
+ min_nb_erb_freqs=params.min_nb_freqs)
+ padded = np.pad(audio, (0, (-len(audio)) % params.hop_size))
+ wet = enhance(model, state, torch.from_numpy(padded).unsqueeze(0),
+ pad=True).squeeze(0).numpy()[:len(audio)]
+ if len(wet) != len(audio) or not np.isfinite(wet).all():
+ raise RuntimeError("DeepFilterNet3 returned invalid audio")
+ audio = audio * (1 - args.strength) + wet * args.strength
+ kept = audio[start - left:min(frames, end + overlap) - left].copy()
+ if previous is not None and args.strength:
+ n = len(previous)
+ fade = np.linspace(0, 1, n, dtype=np.float32)
+ kept[:n] = previous * (1 - fade) + kept[:n] * fade
+ target.write(kept[:end - start])
+ previous = kept[end - start:].copy()
+ report(f"Denoising ({device})", end, frames)
+ if target.tell() != frames:
+ raise RuntimeError("Output frame count does not match input")
+ os.replace(temporary, args.target)
+ temporary = None
+ finally:
+ if temporary is not None:
+ temporary.unlink(missing_ok=True)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/lib/project/src/voiceforge/audio.py b/lib/project/src/voiceforge/audio.py
new file mode 100644
index 0000000..7fee0c2
--- /dev/null
+++ b/lib/project/src/voiceforge/audio.py
@@ -0,0 +1,128 @@
+"""Streaming measurement and FFmpeg execution. No shell interpolation."""
+from collections import deque
+import json
+import math
+import os
+from pathlib import Path
+import selectors
+import signal
+import subprocess
+import time
+
+import numpy as np
+import soundfile as sf
+
+from .setup import terminate_group
+
+
+def ffmpeg_run(ffmpeg: str, args: list[str], stage: str, duration: float,
+ progress) -> str:
+ command = [ffmpeg, "-hide_banner", "-nostdin", "-y", "-nostats",
+ "-progress", "pipe:1"] + args
+ progress(stage, 0, duration or None)
+ diagnostics = deque(maxlen=128)
+ pending = b""
+ completed = 0.0
+ last_advance = time.monotonic()
+ with subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+ start_new_session=True) as process:
+ try:
+ with selectors.DefaultSelector() as selector:
+ for stream in (process.stdout, process.stderr):
+ os.set_blocking(stream.fileno(), False)
+ selector.register(stream, selectors.EVENT_READ)
+ while selector.get_map() or process.poll() is None:
+ for key, _ in selector.select(timeout=0.2):
+ data = os.read(key.fd, 8192)
+ if not data:
+ selector.unregister(key.fileobj)
+ continue
+ if key.fileobj is process.stderr:
+ diagnostics.append(data)
+ else:
+ pending += data
+ while b"\n" in pending:
+ line, pending = pending.split(b"\n", 1)
+ if line.startswith(b"out_time_us="):
+ try:
+ value = max(0.0, int(line.split(b"=", 1)[1]) / 1e6)
+ except ValueError:
+ continue
+ if value > completed:
+ last_advance = time.monotonic()
+ completed = value
+ label = stage
+ if time.monotonic() - last_advance > 60:
+ label += " (no new audio progress for 60s; still waiting)"
+ progress(label, min(completed, duration), duration or None)
+ code = process.wait()
+ output = b"".join(diagnostics).decode(errors="replace")
+ if code:
+ raise RuntimeError(f"{stage} failed (exit {code}):\n{output[-8000:]}")
+ progress(stage, duration, duration or None)
+ return output
+ except BaseException:
+ terminate_group(process)
+ raise
+
+
+def measure(path: Path, progress, stage="Analyzing audio") -> dict:
+ """100ms RMS histogram is bounded in memory, including multi-hour input."""
+ with sf.SoundFile(path) as source:
+ if source.channels not in (1, 2):
+ raise ValueError("Only mono or stereo voice recordings are supported")
+ if not len(source):
+ raise ValueError("The recording is empty")
+ peaks = np.zeros(source.channels)
+ sums = np.zeros(source.channels)
+ squares = np.zeros(source.channels)
+ clipped = 0
+ histogram = np.zeros(121, dtype=np.int64)
+ frames = 0
+ progress(stage, 0, len(source))
+ for audio in source.blocks(blocksize=max(1, source.samplerate // 10),
+ dtype="float64", always_2d=True):
+ if not np.isfinite(audio).all():
+ raise ValueError("Recording contains non-finite samples")
+ peaks = np.maximum(peaks, np.max(np.abs(audio), axis=0))
+ sums += audio.sum(axis=0)
+ squares += (audio * audio).sum(axis=0)
+ clipped += int(np.count_nonzero(np.abs(audio) >= 0.9999))
+ rms = float(np.sqrt(np.mean(audio * audio)))
+ db = 20 * math.log10(max(rms, 1e-6))
+ histogram[int(np.clip(round(db) + 120, 0, 120))] += 1
+ frames += len(audio)
+ progress(stage, frames, len(source))
+ cumulative = histogram.cumsum()
+ def percentile(p):
+ return int(np.searchsorted(cumulative, max(1, math.ceil(cumulative[-1] * p)))) - 120
+ quiet, speech = percentile(0.1), percentile(0.8)
+ if speech - quiet < 15:
+ # Pause-dominated recordings: the 80th percentile of all blocks
+ # sits in room tone, so estimate the level from the loud tail
+ # instead of mistaking sparse active speech for silence.
+ speech = max(speech, percentile(0.98))
+ noise_available = speech - quiet >= 15 and quiet > -110
+ return {"frames": frames, "duration_seconds": frames / source.samplerate,
+ "sample_rate": source.samplerate, "channels": source.channels,
+ "subtype": source.subtype,
+ "sample_peak_dbfs": float(20 * np.log10(max(float(peaks.max()), 1e-12))),
+ "channel_rms_dbfs": (20 * np.log10(np.maximum(np.sqrt(squares / frames), 1e-12))).tolist(),
+ "dc_offset": (sums / frames).tolist(), "near_full_scale_samples": clipped,
+ "speech_level_estimate_dbfs": speech,
+ "quiet_blocks_dbfs": quiet,
+ "noise_floor_estimate_dbfs": quiet if noise_available else None,
+ "noise_floor_confidence": "low (quiet-block heuristic, not speech recognition)" if noise_available else "insufficient room tone"}
+
+
+def loudness(ffmpeg, source, settings, duration, progress, stage="Measuring loudness"):
+ output = ffmpeg_run(ffmpeg, ["-i", str(source), "-map", "0:a:0", "-af",
+ f"loudnorm=I={settings.target_lufs}:TP={settings.true_peak_db}:LRA={settings.loudness_range}:print_format=json",
+ "-f", "null", "-"], stage, duration, progress)
+ start, end = output.rfind("{"), output.rfind("}")
+ if start < 0 or end < start:
+ raise RuntimeError("FFmpeg did not return loudness measurements")
+ raw = json.loads(output[start:end + 1])
+ return {key: (float(value) if math.isfinite(float(value)) else None)
+ for key, value in raw.items() if key != "normalization_type"} | {
+ "normalization_type": raw.get("normalization_type")}
diff --git a/lib/project/src/voiceforge/cli.py b/lib/project/src/voiceforge/cli.py
new file mode 100644
index 0000000..52cd1bd
--- /dev/null
+++ b/lib/project/src/voiceforge/cli.py
@@ -0,0 +1,193 @@
+"""Command-line entry point; configuration errors fail before model downloads."""
+import argparse
+from dataclasses import asdict, fields
+import json
+import math
+import os
+from pathlib import Path
+import subprocess
+import sys
+import tempfile
+import tomllib
+
+from rich.console import Console
+
+from . import __version__
+from .audio import ffmpeg_run, loudness, measure
+from .config import PROFILES, Settings, resolve, toml
+from .pipeline import check_target, process, targets_for
+from .progress import Display
+from .setup import data_dir, ensure_ffmpeg
+
+
+def parser():
+ app = argparse.ArgumentParser(prog="producer", description="Local speech cleanup and podcast mastering. Originals are never overwritten.")
+ app.add_argument("--version", action="version", version=__version__)
+ sub = app.add_subparsers(dest="command", required=True)
+ for name in ("process", "preview", "config", "analyze"):
+ cmd = sub.add_parser(name)
+ cmd.add_argument("--config", type=Path, help="TOML configuration file")
+ cmd.add_argument("--profile", choices=PROFILES, default=None)
+ cmd.add_argument("--set", action="append", default=[], metavar="KEY=VALUE", help="Override any setting using TOML value syntax")
+ for field in fields(Settings):
+ if field.name == "profile":
+ continue
+ default = getattr(Settings(), field.name)
+ option = "--" + field.name.replace("_", "-")
+ if isinstance(default, bool):
+ cmd.add_argument(option, action=argparse.BooleanOptionalAction, default=None)
+ else:
+ cmd.add_argument(option, type=type(default), default=None)
+ cmd.add_argument("--no-denoise", dest="denoiser", action="store_const", const="none")
+ if name == "config":
+ cmd.add_argument("--json", action="store_true")
+ else:
+ cmd.add_argument("inputs", type=Path, nargs="+" if name == "process" else 1)
+ if name in ("process", "preview"):
+ cmd.add_argument("--output-dir", type=Path, default=None,
+ help="Directory for generated outputs (default: beside each input)")
+ cmd.add_argument("--overwrite", action="store_true", help="Replace generated outputs, never input files")
+ if name == "process":
+ cmd.add_argument("-o", "--output", type=Path, help="Explicit WAV destination (one input only)")
+ if name == "preview":
+ cmd.add_argument("--start", type=float, default=0)
+ cmd.add_argument("--duration", type=float, default=30)
+ cmd.add_argument("--profiles", nargs="+", choices=PROFILES, default=["natural", "narrator", "radio"])
+ sub.add_parser("doctor", help="Check tools and driver without downloading AI")
+ setup = sub.add_parser("setup", help="Download and self-test the isolated AI backend")
+ setup.add_argument("--device", choices=("auto", "cpu", "cuda"), default="auto")
+ return app
+
+
+def settings_for(args, profile=None):
+ overrides = {field.name: getattr(args, field.name) for field in fields(Settings)
+ if field.name != "profile" and getattr(args, field.name, None) is not None}
+ for item in args.set:
+ key, sep, value = item.partition("=")
+ if not sep:
+ raise ValueError("--set expects KEY=VALUE")
+ # Strings can be supplied without TOML quotes for CLI convenience.
+ try:
+ parsed = tomllib.loads("value = " + value)["value"]
+ except tomllib.TOMLDecodeError:
+ parsed = value
+ overrides[key.strip()] = parsed
+ return resolve(args.config, profile or args.profile, overrides)
+
+
+def require_input(source: Path) -> None:
+ if not source.is_file() or source.suffix.lower() != ".wav":
+ raise ValueError(f"Input must be an existing WAV file: {source}")
+
+
+def main():
+ app = parser()
+ args = app.parse_args()
+ console = Console(stderr=True)
+ try:
+ if args.command == "doctor":
+ from .ai import _cuda_available
+ ffmpeg = ensure_ffmpeg()
+ version = subprocess.run([ffmpeg, "-version"], capture_output=True, text=True, check=True).stdout.splitlines()[0]
+ console.print(version, markup=False)
+ console.print(f"FFmpeg: {ffmpeg}\nData: {data_dir()}\nCUDA driver available: {_cuda_available()}\nAI is installed and self-tested by './producer.sh setup'.", markup=False)
+ return
+ if args.command == "setup":
+ from .ai import ensure_ai
+ with Display() as display:
+ ensure_ffmpeg()
+ python = ensure_ai(args.device, display.update, verify=True)
+ console.print(f"AI ready: {python}", markup=False)
+ return
+ settings = settings_for(args)
+ if args.command == "config":
+ print(json.dumps(asdict(settings), indent=2) if args.json else toml(settings), end="\n" if args.json else "")
+ return
+ with Display() as display:
+ if args.command == "analyze":
+ stats = measure(args.inputs[0], display.update)
+ stats["loudness"] = loudness(ensure_ffmpeg(), args.inputs[0], settings, stats["duration_seconds"], display.update)
+ print(json.dumps(stats, indent=2, allow_nan=False))
+ return
+ if args.command == "process":
+ if args.output and len(args.inputs) != 1:
+ raise ValueError("--output requires exactly one input")
+ if args.output and args.output_dir:
+ raise ValueError("--output and --output-dir cannot be combined")
+ # Without an explicit destination, outputs are written beside
+ # each input; --output-dir keeps every result in one directory.
+ outputs = [args.output or (args.output_dir or p.parent) / f"{p.stem}.{settings.profile}.wav"
+ for p in args.inputs]
+ # Preflight the whole plan so a predictable conflict cannot
+ # fail a batch after earlier inputs have already been published.
+ sources = {p.resolve() for p in args.inputs}
+ claimed: set[Path] = set()
+ for source, output in zip(args.inputs, outputs):
+ require_input(source)
+ targets = targets_for(output.absolute(), settings)
+ resolved = {target.resolve() for target in targets}
+ if len(resolved) < len(targets) or resolved & claimed:
+ raise ValueError("Inputs have duplicate output names; process them separately with --output")
+ claimed |= resolved
+ if resolved & sources or any(
+ target.exists() and os.path.samefile(source, target) for target in targets):
+ raise ValueError("An output would overwrite an input recording")
+ for target in targets:
+ check_target(source, target, args.overwrite)
+ for source, output in zip(args.inputs, outputs):
+ report = process(source, output, settings, display.update, args.overwrite)
+ console.print(f"Written: {output} | {report['master_loudness']['input_i']} LUFS | {report['master_loudness']['input_tp']} dBTP", markup=False)
+ for warning in report["warnings"]:
+ console.print(f"Warning: {warning}", markup=False)
+ return
+ if not math.isfinite(args.start) or args.start < 0 or not 0 < args.duration <= 300:
+ raise ValueError("Preview start must be >= 0 and duration must be in (0, 300] seconds")
+ if len(set(args.profiles)) != len(args.profiles):
+ raise ValueError("Preview profiles must be unique")
+ if any(item.partition("=")[0].strip() == "profile" for item in args.set):
+ raise ValueError("Preview profiles come from --profiles; remove the profile override from --set")
+ source = args.inputs[0]
+ require_input(source)
+ original = source.resolve()
+ output_dir = args.output_dir or source.parent
+ # Preflight every preview destination before extracting the excerpt.
+ claimed: set[Path] = set()
+ for profile in args.profiles:
+ output = output_dir / f"{source.stem}.{profile}.preview.wav"
+ if output.resolve() == original:
+ raise ValueError("Preview output would overwrite the original")
+ targets = [output, output.with_suffix(".report.json")]
+ resolved = {target.resolve() for target in targets}
+ if len(resolved) < len(targets) or resolved & claimed:
+ raise ValueError("Preview outputs have duplicate names")
+ claimed |= resolved
+ for target in targets:
+ check_target(original, target, args.overwrite)
+ output_dir.mkdir(parents=True, exist_ok=True)
+ with tempfile.TemporaryDirectory(prefix=".preview-", dir=output_dir) as tmp:
+ excerpt = Path(tmp) / "excerpt.wav"
+ ffmpeg_run(ensure_ffmpeg(), ["-ss", str(args.start), "-i", str(original), "-t", str(args.duration),
+ "-map", "0:a:0", "-c:a", "pcm_f32le", str(excerpt)], "Extracting preview", args.duration, display.update)
+ for profile in args.profiles:
+ configured = settings_for(args, profile)
+ # Equal loudness prevents a louder preset winning an unfair A/B.
+ configured.normalize = True
+ configured.limiter = True
+ output = output_dir / f"{source.stem}.{profile}.preview.wav"
+ report = process(excerpt, output, configured, display.update, args.overwrite,
+ provenance={"original": str(original), "preview_start_seconds": args.start,
+ "requested_duration_seconds": args.duration,
+ "normalization_forced": True})
+ console.print(f"Preview: {output} | {report['master_loudness']['input_i']} LUFS | {report['master_loudness']['input_tp']} dBTP", markup=False)
+ for warning in report["warnings"]:
+ console.print(f"Warning: {warning}", markup=False)
+ except KeyboardInterrupt:
+ console.print("Cancelled; temporary files removed. Completed outputs may remain.")
+ raise SystemExit(130)
+ except (ValueError, OSError, RuntimeError, subprocess.SubprocessError) as error:
+ console.print(f"Error: {error}", markup=False)
+ raise SystemExit(1)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/lib/project/src/voiceforge/config.py b/lib/project/src/voiceforge/config.py
new file mode 100644
index 0000000..9cbd59d
--- /dev/null
+++ b/lib/project/src/voiceforge/config.py
@@ -0,0 +1,118 @@
+"""Flat, strict TOML configuration; CLI overrides are applied last."""
+from dataclasses import asdict, dataclass, fields
+import math
+from pathlib import Path
+import tomllib
+
+
+@dataclass
+class Settings:
+ profile: str = "natural"
+ denoiser: str = "deepfilter"
+ device: str = "auto"
+ denoise_strength: float = 0.85
+ fft_reduction_db: float = 8.0
+ channel: str = "auto"
+ highpass: bool = True
+ highpass_hz: float = 70.0
+ lowpass: bool = False
+ lowpass_hz: float = 16000.0
+ hum_hz: int = 0
+ leveling: bool = True
+ level_target_db: float = -23.0
+ max_gain_db: float = 18.0
+ expansion: bool = False
+ expansion_threshold_db: float = -50.0
+ expansion_ratio: float = 1.5
+ expansion_range_db: float = 12.0
+ eq: bool = True
+ warmth_db: float = 0.5
+ mud_db: float = -1.5
+ presence_db: float = 1.0
+ compression: bool = True
+ compressor_threshold_db: float = -21.0
+ compressor_ratio: float = 2.0
+ compressor_attack_ms: float = 15.0
+ compressor_release_ms: float = 150.0
+ deess: bool = True
+ deess_intensity: float = 0.15
+ deess_amount: float = 0.4
+ normalize: bool = True
+ target_lufs: float = -19.0
+ loudness_range: float = 9.0
+ limiter: bool = True
+ true_peak_db: float = -1.5
+ sample_rate: int = 48000
+ mp3: bool = False
+ mp3_bitrate: int = 192
+
+
+PROFILES = {
+ "natural": {},
+ "narrator": {"compressor_ratio": 1.6, "compressor_attack_ms": 25.0,
+ "presence_db": 0.5, "warmth_db": 1.0, "loudness_range": 11.0},
+ "radio": {"compressor_ratio": 3.5, "compressor_threshold_db": -24.0,
+ "compressor_attack_ms": 8.0, "compressor_release_ms": 100.0,
+ "warmth_db": 2.0, "mud_db": -2.0, "presence_db": 2.0,
+ "deess_intensity": 0.25, "loudness_range": 6.0},
+ "cleanup-only": {"leveling": False, "eq": False, "compression": False,
+ "deess": False, "normalize": False, "limiter": False},
+}
+
+RANGES = {
+ "denoise_strength": (0, 1), "fft_reduction_db": (0.01, 30),
+ "highpass_hz": (20, 300), "lowpass_hz": (4000, 22000),
+ "level_target_db": (-40, -12), "max_gain_db": (0, 30),
+ "expansion_threshold_db": (-90, -20), "expansion_ratio": (1, 4),
+ "expansion_range_db": (0, 40), "warmth_db": (-12, 12),
+ "mud_db": (-12, 12), "presence_db": (-12, 12),
+ "compressor_threshold_db": (-50, -5), "compressor_ratio": (1, 10),
+ "compressor_attack_ms": (0.1, 200), "compressor_release_ms": (10, 2000),
+ "deess_intensity": (0, 1), "deess_amount": (0, 1),
+ "target_lufs": (-30, -12), "loudness_range": (1, 20),
+ "true_peak_db": (-9, -0.5),
+}
+
+
+def resolve(path: Path | None = None, profile: str | None = None,
+ overrides: dict | None = None) -> Settings:
+ values = tomllib.loads(path.read_text()) if path else {}
+ chosen = (overrides or {}).get("profile", profile or values.get("profile", "natural"))
+ if not isinstance(chosen, str) or chosen not in PROFILES:
+ raise ValueError(f"Unknown profile {chosen!r}; choose {', '.join(PROFILES)}")
+ merged = asdict(Settings()) | PROFILES[chosen] | values | {"profile": chosen}
+ merged.update(overrides or {})
+ defaults = asdict(Settings())
+ unknown = merged.keys() - defaults.keys()
+ if unknown:
+ raise ValueError(f"Unknown setting(s): {', '.join(sorted(unknown))}")
+ for key, value in merged.items():
+ expected = type(defaults[key])
+ if expected is float:
+ if type(value) not in (int, float) or not math.isfinite(value):
+ raise ValueError(f"{key} must be a finite number")
+ elif type(value) is not expected:
+ raise ValueError(f"{key} must be {expected.__name__}")
+ if key in RANGES and not RANGES[key][0] <= value <= RANGES[key][1]:
+ raise ValueError(f"{key} must be between {RANGES[key][0]} and {RANGES[key][1]}")
+ for key, options in {
+ "denoiser": ("deepfilter", "fft", "none"), "device": ("auto", "cpu", "cuda"),
+ "channel": ("auto", "left", "right", "mix"), "hum_hz": (0, 50, 60),
+ "sample_rate": (44100, 48000), "mp3_bitrate": (128, 160, 192, 224, 256, 320),
+ }.items():
+ if merged[key] not in options:
+ raise ValueError(f"{key} must be one of {options}")
+ if merged["lowpass"] and merged["lowpass_hz"] >= merged["sample_rate"] / 2:
+ raise ValueError("lowpass_hz must be below the output Nyquist frequency")
+ if merged["normalize"] and not merged["limiter"]:
+ raise ValueError("Normalization includes true-peak limiting; use --no-normalize to bypass it")
+ return Settings(**merged)
+
+
+def toml(settings: Settings) -> str:
+ lines = ["# VoiceForge configuration. CLI flags override these values."]
+ for field in fields(settings):
+ value = getattr(settings, field.name)
+ text = str(value).lower() if isinstance(value, bool) else repr(value)
+ lines.append(f"{field.name} = {text}")
+ return "\n".join(lines) + "\n"
diff --git a/lib/project/src/voiceforge/pipeline.py b/lib/project/src/voiceforge/pipeline.py
new file mode 100644
index 0000000..a143dcf
--- /dev/null
+++ b/lib/project/src/voiceforge/pipeline.py
@@ -0,0 +1,199 @@
+"""Offline mono voice mastering with float intermediates and atomic outputs."""
+from dataclasses import asdict
+import fcntl
+import json
+import math
+import os
+from pathlib import Path
+import tempfile
+
+from . import __version__
+from .audio import ffmpeg_run, loudness, measure
+from .config import Settings
+from .setup import ensure_ffmpeg
+
+
+def filters(settings: Settings) -> list[str]:
+ s = settings
+ result = []
+ if s.expansion:
+ result.append(f"agate=threshold={10 ** (s.expansion_threshold_db / 20)}:ratio={s.expansion_ratio}:range={10 ** (-s.expansion_range_db / 20)}:attack=10:release=250:detection=rms")
+ if s.eq:
+ result += [f"equalizer=f=140:t=q:w=0.7:g={s.warmth_db}",
+ f"equalizer=f=300:t=q:w=0.8:g={s.mud_db}",
+ f"equalizer=f=3500:t=q:w=0.7:g={s.presence_db}"]
+ if s.compression:
+ result.append(f"acompressor=threshold={10 ** (s.compressor_threshold_db / 20)}:ratio={s.compressor_ratio}:attack={s.compressor_attack_ms}:release={s.compressor_release_ms}:knee=4:makeup=1:detection=rms")
+ if s.deess:
+ result.append(f"deesser=i={s.deess_intensity}:m={s.deess_amount}:f=0.5")
+ if s.lowpass:
+ result.append(f"lowpass=f={s.lowpass_hz}:p=2")
+ return result or ["anull"]
+
+
+def targets_for(destination: Path, settings: Settings) -> list[Path]:
+ """Every path a successful run publishes: master WAV, report, optional MP3."""
+ targets = [destination, destination.with_suffix(".report.json")]
+ if settings.mp3:
+ targets.append(destination.with_suffix(".mp3"))
+ return targets
+
+
+def check_target(source: Path, target: Path, overwrite: bool) -> None:
+ """Preflight one published path; the original recording is never touched."""
+ if target.resolve() == source or (target.exists() and os.path.samefile(source, target)):
+ raise ValueError("Refusing to overwrite the original recording")
+ # lexists also rejects dangling symlinks, which publication would only hit late.
+ if os.path.lexists(target) and not overwrite:
+ raise FileExistsError(f"Output exists: {target}; use --overwrite to replace outputs")
+ if target.exists() and not target.is_file():
+ raise ValueError(f"Output is not a regular file: {target}")
+
+
+def publish(source: Path, target: Path, overwrite: bool):
+ if overwrite:
+ os.replace(source, target)
+ else:
+ # Link is atomic and fails if another process created the destination.
+ os.link(source, target)
+ source.unlink()
+
+
+def process(source: Path, destination: Path, settings: Settings, progress,
+ overwrite=False, provenance: dict | None = None) -> dict:
+ source, destination = source.resolve(), destination.absolute()
+ if not source.is_file() or source.suffix.lower() != ".wav":
+ raise ValueError(f"Input must be an existing WAV file: {source}")
+ if destination.suffix.lower() != ".wav":
+ raise ValueError("Master output must have a .wav extension")
+ report_path = destination.with_suffix(".report.json")
+ for target in targets_for(destination, settings):
+ check_target(source, target, overwrite)
+ ffmpeg = ensure_ffmpeg()
+ original = measure(source, progress, "Analyzing original")
+ if original["sample_peak_dbfs"] < -100:
+ raise ValueError("Recording is silent or too quiet to master safely")
+ channel = settings.channel
+ if original["channels"] == 2 and channel == "auto":
+ raise ValueError("Stereo input: select --channel left, right, or mix explicitly. Auto downmix could cancel your voice.")
+ if original["channels"] == 1 and channel == "right":
+ raise ValueError("Cannot select the right channel of a mono recording")
+ duration = original["duration_seconds"]
+ warnings = []
+ if original["near_full_scale_samples"]:
+ warnings.append("Input has near-full-scale samples: inspect for clipping; lost peaks cannot be reliably restored.")
+ if max(abs(x) for x in original["dc_offset"]) > 0.01:
+ warnings.append("Significant input DC offset detected; enable highpass filtering to remove it.")
+ destination.parent.mkdir(parents=True, exist_ok=True)
+ with tempfile.TemporaryDirectory(prefix=".voiceforge-", dir=destination.parent) as temporary:
+ work = Path(temporary)
+ prepared, cleaned, shaped = [work / name for name in ("prepared.wav", "cleaned.wav", "shaped.wav")]
+ input_filters = []
+ if original["channels"] == 2:
+ input_filters.append({"left": "pan=mono|c0=c0", "right": "pan=mono|c0=c1",
+ "mix": "pan=mono|c0=0.5*c0+0.5*c1"}[channel])
+ if settings.highpass:
+ input_filters.append(f"highpass=f={settings.highpass_hz}:p=2")
+ if settings.hum_hz:
+ input_filters += [f"equalizer=f={settings.hum_hz * n}:t=q:w=25:g=-18" for n in (1, 2, 3)]
+ ffmpeg_run(ffmpeg, ["-i", str(source), "-map", "0:a:0", "-af",
+ ",".join(input_filters or ["anull"]), "-ar", "48000", "-ac", "1",
+ "-c:a", "pcm_f32le", "-rf64", "auto", str(prepared)], "Preparing mono 48 kHz audio", duration, progress)
+ if settings.denoiser == "deepfilter" and settings.denoise_strength > 0:
+ from .ai import denoise
+ denoise(prepared, cleaned, settings.device, settings.denoise_strength, progress)
+ elif settings.denoiser == "fft" and settings.denoise_strength > 0:
+ ffmpeg_run(ffmpeg, ["-i", str(prepared), "-af",
+ f"afftdn=nr={max(0.01, settings.fft_reduction_db * settings.denoise_strength)}:tn=1",
+ "-c:a", "pcm_f32le", "-rf64", "auto", str(cleaned)], "FFT noise reduction", duration, progress)
+ else:
+ cleaned = prepared
+ clean_stats = measure(cleaned, progress, "Measuring cleaned voice")
+ if clean_stats["sample_peak_dbfs"] < -100:
+ raise ValueError("The selected audio is silent or too quiet to master safely; "
+ "check --channel, the mix, and cleanup settings")
+ gain = 0.0
+ if settings.leveling:
+ speech_db = clean_stats["speech_level_estimate_dbfs"]
+ if speech_db <= -110:
+ warnings.append("Speech level estimate is unreliable; leveling skipped. "
+ "Inspect levels or rely on normalization.")
+ else:
+ desired = settings.level_target_db - speech_db
+ gain = max(-settings.max_gain_db, min(settings.max_gain_db, desired))
+ if abs(desired) > settings.max_gain_db:
+ warnings.append("Input level correction reached its configured gain limit.")
+ chain = ([f"volume={gain}dB"] if gain else []) + filters(settings)
+ ffmpeg_run(ffmpeg, ["-i", str(cleaned), "-af", ",".join(chain),
+ "-c:a", "pcm_f32le", "-rf64", "auto", str(shaped)], "EQ, dynamics and de-essing", duration, progress)
+ before = loudness(ffmpeg, shaped, settings, duration, progress, "Loudness analysis (pass 1)")
+ if settings.normalize and any(before.get(k) is None for k in ("input_i", "input_tp", "input_lra", "input_thresh", "target_offset")):
+ raise ValueError("Cannot normalize: insufficient measurable audio after processing")
+ final_filters = []
+ if settings.normalize:
+ final_filters.append(f"loudnorm=I={settings.target_lufs}:TP={settings.true_peak_db}:LRA={settings.loudness_range}:"
+ f"measured_I={before['input_i']}:measured_TP={before['input_tp']}:measured_LRA={before['input_lra']}:"
+ f"measured_thresh={before['input_thresh']}:offset={before['target_offset']}:linear=true:print_format=json")
+ elif settings.limiter:
+ final_filters += ["aresample=192000", f"alimiter=limit={10 ** (settings.true_peak_db / 20)}:level=false:latency=true"]
+ final_filters.append(f"aresample={settings.sample_rate}:output_sample_bits=24:dither_method=triangular")
+ master = work / "master.wav"
+ render_log = ffmpeg_run(ffmpeg, ["-i", str(shaped), "-af", ",".join(final_filters),
+ "-ar", str(settings.sample_rate), "-c:a", "pcm_s24le", "-rf64", "auto", str(master)],
+ "Mastering (pass 2)", duration, progress)
+ # Bounded disk use: drop consumed intermediates instead of waiting for
+ # the temporary directory teardown at the end of the run.
+ shaped.unlink(missing_ok=True)
+ for intermediate in {prepared, cleaned}:
+ intermediate.unlink(missing_ok=True)
+ normalization_mode = "disabled"
+ if settings.normalize:
+ start, end = render_log.rfind("{"), render_log.rfind("}")
+ if start < 0 or end < start:
+ raise RuntimeError("FFmpeg did not report the applied normalization mode")
+ normalization_mode = json.loads(render_log[start:end + 1])["normalization_type"]
+ final = loudness(ffmpeg, master, settings, duration, progress, "Verifying exported master")
+ final_stats = measure(master, progress, "Checking exported samples")
+ if abs(final_stats["duration_seconds"] - duration) > 0.02:
+ raise RuntimeError("Output duration differs from input by more than 20 ms")
+ if (final_stats["channels"] != 1 or final_stats["sample_rate"] != settings.sample_rate
+ or final_stats["subtype"] != "PCM_24"):
+ raise RuntimeError("Exported master is not the expected mono PCM_24 WAV")
+ if settings.normalize and (final["input_i"] is None or abs(final["input_i"] - settings.target_lufs) > 0.5):
+ warnings.append("Master misses the loudness target by more than 0.5 LU; inspect the report before publishing.")
+ if settings.limiter and final["input_tp"] is not None and final["input_tp"] > settings.true_peak_db + 0.1:
+ warnings.append("Master exceeds the true-peak target by more than 0.1 dB; inspect before publishing.")
+ if final_stats["near_full_scale_samples"]:
+ warnings.append("Export has near-full-scale samples. Enable limiting or lower gains.")
+ mp3_stats = None
+ if settings.mp3:
+ mp3 = work / "delivery.mp3"
+ ffmpeg_run(ffmpeg, ["-i", str(master), "-c:a", "libmp3lame", "-b:a",
+ f"{settings.mp3_bitrate}k", str(mp3)], "Encoding MP3", duration, progress)
+ mp3_stats = loudness(ffmpeg, mp3, settings, duration, progress, "Verifying decoded MP3")
+ if mp3_stats["input_tp"] is not None and mp3_stats["input_tp"] > settings.true_peak_db + 0.1:
+ warnings.append("MP3 encoding increased true peak beyond the configured ceiling; use more headroom.")
+ report = {"voiceforge_version": __version__, "source": str(source),
+ "output": str(destination), "settings": asdict(settings), "input": original,
+ "cleaned": clean_stats, "level_correction_db": gain,
+ "pre_master_loudness": before, "master_loudness": final,
+ "master": final_stats, "mp3_loudness": mp3_stats,
+ "normalization_mode": normalization_mode,
+ "warnings": warnings}
+ if provenance:
+ report["provenance"] = provenance
+ staged_report = work / "report.json"
+ staged_report.write_text(json.dumps(report, indent=2, allow_nan=False) + "\n")
+ # Publish only after all processing and verification succeeded. The
+ # directory lock serializes publication so concurrent jobs targeting
+ # the same output directory cannot interleave WAV/report/MP3 sets.
+ publication_fd = os.open(destination.parent, os.O_RDONLY)
+ try:
+ fcntl.flock(publication_fd, fcntl.LOCK_EX)
+ if settings.mp3:
+ publish(mp3, destination.with_suffix(".mp3"), overwrite)
+ publish(staged_report, report_path, overwrite)
+ publish(master, destination, overwrite)
+ finally:
+ os.close(publication_fd)
+ return report
diff --git a/lib/project/src/voiceforge/progress.py b/lib/project/src/voiceforge/progress.py
new file mode 100644
index 0000000..a8c5476
--- /dev/null
+++ b/lib/project/src/voiceforge/progress.py
@@ -0,0 +1,40 @@
+"""A refresh thread keeps elapsed time moving even during native processing."""
+import time
+
+from rich.console import Console
+from rich.progress import (BarColumn, Progress, SpinnerColumn, TaskProgressColumn,
+ TextColumn, TimeElapsedColumn, TimeRemainingColumn)
+
+
+class Display:
+ def __init__(self):
+ self.console = Console(stderr=True)
+ self.progress = Progress(SpinnerColumn(), TextColumn("{task.description}"),
+ BarColumn(), TaskProgressColumn(), TimeElapsedColumn(),
+ TimeRemainingColumn(), console=self.console,
+ refresh_per_second=5, disable=not self.console.is_terminal)
+ self.task = None
+ self.stage = ""
+ self.started = time.monotonic()
+ self.last_print = 0.0
+
+ def __enter__(self):
+ self.progress.start()
+ return self
+
+ def __exit__(self, *args):
+ self.progress.stop()
+
+ def update(self, stage, completed=None, total=None):
+ changed = stage != self.stage
+ if changed:
+ if self.task is not None:
+ self.progress.remove_task(self.task)
+ self.task = self.progress.add_task(stage, total=total)
+ self.stage = stage
+ self.progress.update(self.task, completed=completed if total else 0, total=total)
+ now = time.monotonic()
+ if not self.console.is_terminal and (changed or now - self.last_print >= 5):
+ fraction = f" {min(100, 100 * (completed or 0) / total):.0f}%" if total else ""
+ self.console.print(f"[{now - self.started:7.1f}s] {stage}{fraction}", markup=False)
+ self.last_print = now
diff --git a/lib/project/src/voiceforge/setup.py b/lib/project/src/voiceforge/setup.py
new file mode 100644
index 0000000..9afeb44
--- /dev/null
+++ b/lib/project/src/voiceforge/setup.py
@@ -0,0 +1,203 @@
+"""User-local tools. No sudo and no changes to the application's interpreter."""
+from __future__ import annotations
+
+from collections import deque
+from collections.abc import Callable
+import json
+import os
+from pathlib import Path
+import selectors
+import shutil
+import signal
+import subprocess
+import sys
+import time
+
+Progress = Callable[[str, float | None, float | None], None]
+
+
+def data_dir() -> Path:
+ """The VoiceForge data/prefix directory.
+
+ VOICEFORGE_HOME wins (the single-directory launcher exports it), then
+ XDG_DATA_HOME/voiceforge, then ~/.local/share/voiceforge. A relative
+ VOICEFORGE_HOME resolves against the current directory.
+ """
+ home = os.environ.get("VOICEFORGE_HOME")
+ if home:
+ return Path(home).absolute()
+ root = Path(os.environ.get("XDG_DATA_HOME", str(Path.home() / ".local/share")))
+ if not root.is_absolute():
+ root = Path.home() / ".local/share"
+ return root / "voiceforge"
+
+
+def tool_env() -> dict[str, str]:
+ """Sanitized environment for bootstrap, setup, and worker subprocesses.
+
+ Host Python settings are removed so a virtualenv or sitecustomize on the
+ machine cannot leak into managed environments. Installation-target,
+ index, and constraint overrides are dropped and package-manager
+ configuration files disabled, so packages can only come from the
+ explicitly passed trusted indexes and land inside the data directory;
+ every cache (pip, uv, managed Python, XDG, temporary files) stays there
+ too.
+ """
+ env = os.environ.copy()
+ for name in ("PYTHONPATH", "PYTHONHOME", "VIRTUAL_ENV",
+ "PIP_TARGET", "PIP_PREFIX", "PIP_USER", "PIP_FIND_LINKS",
+ "PIP_INDEX_URL", "PIP_EXTRA_INDEX_URL", "PIP_CONSTRAINT",
+ "UV_INDEX_URL", "UV_DEFAULT_INDEX", "UV_EXTRA_INDEX_URL",
+ "UV_INDEX", "UV_FIND_LINKS", "UV_CONSTRAINT"):
+ env.pop(name, None)
+ env["PYTHONNOUSERSITE"] = "1"
+ env["PIP_CONFIG_FILE"] = os.devnull
+ env["UV_NO_CONFIG"] = "1"
+ cache = data_dir() / "cache"
+ env["XDG_CACHE_HOME"] = str(cache)
+ env["PIP_CACHE_DIR"] = str(cache / "pip")
+ env["UV_CACHE_DIR"] = str(cache / "uv")
+ env["UV_PYTHON_INSTALL_DIR"] = str(data_dir() / "uv/python")
+ tmp = data_dir() / "tmp"
+ try:
+ tmp.mkdir(parents=True, exist_ok=True)
+ env["TMPDIR"] = str(tmp)
+ except OSError:
+ pass # Best effort: never break subprocesses over temporary-file containment.
+ return env
+
+
+def terminate_group(process: subprocess.Popen) -> None:
+ """SIGTERM the child's whole process group, escalate to SIGKILL, and reap.
+
+ The group is signalled even when the leader has already exited, because
+ descendants can outlive it while still holding the output pipe open.
+ A race between exit and signalling is tolerated.
+ """
+ for sig in (signal.SIGTERM, signal.SIGKILL):
+ try:
+ os.killpg(process.pid, sig)
+ except ProcessLookupError:
+ pass
+ try:
+ process.wait(timeout=5)
+ except subprocess.TimeoutExpired:
+ continue
+ try:
+ os.killpg(process.pid, 0)
+ except ProcessLookupError:
+ return
+ process.wait()
+
+
+def run_process(args: list[str], stage: str, progress: Progress | None = None,
+ *, env: dict[str, str] | None = None) -> None:
+ """Drain output without blocking; send heartbeat callbacks every 0.25s.
+
+ Worker JSON records carry frame/byte counts. Other processes report elapsed seconds
+ as completed with total=None. Only the last 64 KiB of diagnostics are retained.
+ Callback exceptions (including cancellation) terminate the process group.
+ """
+ started = time.monotonic()
+ current_stage, completed, total = stage, None, None
+ diagnostics: deque[bytes] = deque(maxlen=16)
+ pending = b""
+ if progress:
+ progress(stage, None, None)
+ with subprocess.Popen(args, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
+ env=env, start_new_session=True) as process:
+ assert process.stdout is not None
+ os.set_blocking(process.stdout.fileno(), False)
+ try:
+ with selectors.DefaultSelector() as selector:
+ selector.register(process.stdout, selectors.EVENT_READ)
+ while selector.get_map() or process.poll() is None:
+ for key, _ in selector.select(timeout=0.25):
+ chunk = os.read(key.fd, 4096)
+ if not chunk:
+ selector.unregister(key.fileobj)
+ continue
+ diagnostics.append(chunk)
+ pending += chunk
+ while b"\n" in pending:
+ line, pending = pending.split(b"\n", 1)
+ if line.startswith(b"VOICEFORGE_PROGRESS "):
+ event = json.loads(line[len(b"VOICEFORGE_PROGRESS "):])
+ current_stage, completed, total = event
+ if len(pending) > 65536:
+ pending = pending[-4096:]
+ if progress:
+ progress(current_stage, completed if total is not None else
+ time.monotonic() - started, total)
+ code = process.wait()
+ if code:
+ detail = b"".join(diagnostics).decode("utf-8", errors="replace")
+ raise RuntimeError(f"{stage} failed (exit {code}):\n{detail}")
+ except BaseException:
+ terminate_group(process)
+ raise
+
+
+def _uv_supports_relocatable(uv: str) -> bool:
+ """Whether this uv understands the venv flags AI setup relies on.
+
+ A host uv can be older than the pinned one; capability is checked
+ directly instead of trusting the version string.
+ """
+ try:
+ result = subprocess.run([uv, "venv", "--help"], capture_output=True,
+ text=True, check=True)
+ except (OSError, subprocess.SubprocessError):
+ return False
+ return "--relocatable" in result.stdout + result.stderr
+
+
+def ensure_uv(progress: Progress | None = None) -> str:
+ """Return a uv that supports the flags AI setup relies on.
+
+ The launcher's contained bootstrap uv is preferred; a host uv is trusted
+ only if it actually supports --relocatable. Otherwise uv==0.8.17 is
+ installed from PyPI (trust PyPI's distribution, not a downloaded shell
+ script).
+ """
+ contained = data_dir() / "bootstrap/bin/uv"
+ if contained.is_file() and os.access(contained, os.X_OK):
+ return str(contained)
+ found = shutil.which("uv")
+ if found and _uv_supports_relocatable(found):
+ return found
+ root = contained.parent.parent
+ root.parent.mkdir(parents=True, exist_ok=True)
+ python = sys.executable or shutil.which("python3")
+ env = tool_env()
+ run_process([python, "-m", "venv", str(root)], "Creating uv bootstrap",
+ progress, env=env)
+ run_process([str(root / "bin/python"), "-m", "pip", "install",
+ "--index-url", "https://pypi.org/simple", "uv==0.8.17"],
+ "Installing uv from PyPI", progress, env=env)
+ if not (contained.is_file() and os.access(contained, os.X_OK)):
+ raise RuntimeError("Installing uv from PyPI did not produce a usable "
+ "uv; check for host pip settings (such as "
+ "PIP_TARGET) that redirect installations outside "
+ "the data directory.")
+ return str(contained)
+
+
+def ensure_ffmpeg() -> str:
+ """Return the bundled FFmpeg, or a system one where the bundle is
+ absent, not ffprobe.
+
+ The launcher installs imageio-ffmpeg via the bundled-ffmpeg extra; direct
+ package installs should declare it as a dependency. This helper never
+ runs pip in the current interpreter.
+ """
+ try:
+ import imageio_ffmpeg
+ return imageio_ffmpeg.get_ffmpeg_exe()
+ except (ImportError, RuntimeError) as error:
+ found = shutil.which("ffmpeg")
+ if found:
+ return found
+ raise RuntimeError("FFmpeg is unavailable. Rebuild the contained "
+ "runtime with './producer.sh --rebuild', or "
+ "install a system FFmpeg.") from error
generated by cgit v1.2.3 (git 2.25.1) at 2026-09-08 09:29:56 +0000