fix: getting rnnoise model

2026-07-04 15:43:17 +02:00 · 2025-10-12 18:57:55 +02:00 · 2025-10-12 18:57:55 +02:00 · 158a1c7b01
commit 158a1c7b01
parent 06c8121c79
4 changed files with 814 additions and 0 deletions
--- a/README_clean_audio.md
+++ b/README_clean_audio.md
@ -0,0 +1,104 @@
+# clean_audio.sh — automatic speech cleaning (FFmpeg)
+
+This script batch‑cleans noisy speech recordings with ffmpeg using simple, reliable filters tuned for ASR (e.g., faster‑whisper). By default it REQUIRES RNNoise (arnndn) and will try to auto‑discover or download a model. You can opt‑in to fallback filters with `--allow-fallback`.
+
+## Install
+
+- Required: ffmpeg. Most distros: `sudo pacman -S ffmpeg` or `sudo apt install ffmpeg`.
+- Recommended: ffmpeg with `arnndn` filter and an RNNoise model file (e.g., from Mozilla RNNoise community models). The script will auto-detect common model locations or download one via `Bash/get_rnnoise_model.sh`. You can pass a specific model with `-m /path/to/model.nn`.
+
+Make executable:
+
+```bash
+chmod +x Bash/clean_audio.sh
+```
+
+## Quick start
+
+- Single file, default ASR preset (16k mono, denoise, high‑pass, limiter):
+```bash
+Bash/clean_audio.sh path/to/file.wav
+```
+This produces `path/to/file_clean.wav`.
+
+- Whole folder, 4 parallel jobs, output to `cleaned/`:
+```bash
+Bash/clean_audio.sh path/to/folder -O cleaned -j 4
+```
+
+- Use an RNNoise model explicitly (if your ffmpeg has arnndn):
+```bash
+Bash/clean_audio.sh input.wav -m models/rnnoise_model.nn
+```
+If you omit `-m`, the script will look in common locations; if not found, it will attempt a download via `Bash/get_rnnoise_model.sh`.
+
+Advanced options and compatibility:
+- The cleaner requires RNNoise by default. To allow non-ML fallback filters (afftdn), add `--allow-fallback`.
+- The script uses advanced filter settings when available (e.g., afftdn with `md`). If your ffmpeg build lacks these options, it will error with guidance. Add `--no-advanced` (or `--compat`) to avoid such params.
+
+- Podcast preset (adds dynamics and loudness leveling):
+```bash
+Bash/clean_audio.sh input.wav --preset podcast
+```
+
+## Options
+
+```text
+Usage: clean_audio.sh <input-file|input-dir> [options]
+
+Options:
+  -O, --out-dir DIR         Output directory (default: alongside input file).
+  -e, --ext EXT             Output extension/container: wav|flac (default: wav).
+  -m, --model PATH          RNNoise model file for arnndn; falls back to afftdn if unavailable.
+      --no-ml               Do not use arnndn even if model is provided; use afftdn.
+      --preset NAME         asr (default) | podcast | aggressive
+  -j, --jobs N              Parallel jobs for directory mode (default: 1).
+  -f, --force               Overwrite outputs if they exist.
+  -q, --quiet               Reduce ffmpeg logging noise.
+      --lowpass FREQ        Optional low-pass cutoff (e.g., 8000). Disabled by default.
+      --suffix SUF          Suffix for output basename (default: _clean).
+```
+
+## Designed for ASR (faster‑whisper)
+
+Default output format is mono, 16 kHz, PCM 16‑bit WAV—ideal for most Whisper/faster‑whisper pipelines. You can feed the cleaned files directly into your transcription step.
+
+If you prefer FLAC to save space without quality loss:
+```bash
+Bash/clean_audio.sh input.wav -e flac -O cleaned
+```
+
+## Presets
+
+- asr (default): light, ASR‑friendly cleanup; prevents clipping.
+- podcast: adds gentle dynamics and approximate loudness normalization (single‑pass `loudnorm`).
+- aggressive: heavier gate/dynamics; can suppress background more, but may slightly hurt ASR accuracy—use sparingly.
+
+## Tips
+
+- If you see artifacts from RNNoise, try without a model (uses `afftdn`), or add a low‑pass (e.g., `--lowpass 8000`).
+- For extremely boomy bar recordings, raise high‑pass by editing `HIGHPASS` in the script or add `--lowpass`.
+- If your ffmpeg lacks `arnndn`, you can install a newer build or keep the fallback (afftdn works fine for many cases).
+        - If your ffmpeg is missing features, you can use the helper:
+```bash
+chmod +x Bash/install_ffmpeg_with_arnndn.sh
+Bash/install_ffmpeg_with_arnndn.sh
+```
+It will suggest distro options or build FFmpeg from source with `--enable-librnnoise`.
+
+    RNNoise model downloader helper:
+    ```bash
+    chmod +x Bash/get_rnnoise_model.sh
+    Bash/get_rnnoise_model.sh --yes
+    ```
+    This saves a model into `Bash/models/` which the cleaner will auto-discover.
+
+## Troubleshooting
+
+- “arnndn not available”: Your ffmpeg wasn’t built with it. The script will use `afftdn` instead.
+- Output sounds thin: lower the high‑pass (edit `HIGHPASS=80` in script to `60`) or remove low‑pass.
+- Level too low/high: choose the `podcast` preset for auto leveling, or add your own `loudnorm` in post.
+
+## License
+
+This helper script is provided under the repository’s LICENSE.
--- a/clean_audio.sh
+++ b/clean_audio.sh
@ -0,0 +1,390 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# clean_audio.sh — Fully automatic audio cleaner for speech (ASR-friendly)
+#
+# - Default preset is tuned for ASR (faster-whisper):
+#   mono, 16 kHz, high-pass filter, denoise (RNNoise arnndn by default if model found/provided; else afftdn),
+#   peak limiting to -0.5 dBFS. No aggressive gating/compression by default.
+# - Optional "podcast" preset adds gentle dynamics and loudness leveling.
+# - Accepts single files or directories (recursively).
+# - Optional parallel processing.
+#
+# Dependencies: ffmpeg (arnndn filter recommended for best results)
+# Optional: an RNNoise model file for arnndn (auto-discovered if present; otherwise falls back to afftdn)
+#
+# Usage examples:
+#   Bash/clean_audio.sh input.wav                      # -> input_clean.wav (same folder)
+#   Bash/clean_audio.sh input.wav -O out_dir           # -> out_dir/input_clean.wav
+#   Bash/clean_audio.sh input_dir -O cleaned/ -j 4     # -> processes all audio files in dir
+#   Bash/clean_audio.sh input.wav -m models/rn.nn      # -> use RNNoise model
+#   Bash/clean_audio.sh input.wav --preset podcast     # -> add dynamics leveler
+#
+
+SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)
+
+print_usage() {
+  cat <<EOF
+Usage: $0 <input-file|input-dir> [options]
+
+Options:
+  -O, --out-dir DIR         Output directory (default: alongside input file).
+  -e, --ext EXT             Output extension/container: wav|flac (default: wav).
+  -m, --model PATH          RNNoise model file for arnndn; required by default unless --allow-fallback.
+    --no-ml               Do not use arnndn even if model is provided (requires --allow-fallback).
+      --preset NAME         asr (default) | podcast | aggressive
+  -j, --jobs N              Parallel jobs for directory mode (default: 1).
+  -f, --force               Overwrite outputs if they exist (ffmpeg -y).
+  -q, --quiet               Reduce ffmpeg logging noise.
+      --lowpass FREQ        Optional low-pass cutoff (e.g., 8000). Disabled by default.
+      --suffix SUF          Suffix for output basename (default: _clean).
+  -h, --help                Show this help.
+
+Notes:
+  - Default sample rate is 16 kHz mono PCM 16-bit (good for most speech ASR models).
+  - If arnndn (RNNoise) is used, it usually outperforms afftdn for speech denoise.
+  - The 'podcast' preset adds gentle dynamics and loudness normalization (single-pass).
+EOF
+}
+
+require_cmd() {
+  command -v "$1" >/dev/null 2>&1 || {
+    echo "Error: Required command '$1' not found in PATH" >&2
+    exit 1
+  }
+}
+
+# Defaults
+OUT_DIR=""
+OUT_EXT="wav"
+RN_MODEL=""
+NO_ML=false
+REQUIRE_ML=true        # default: require RNNoise; install/guide if missing; fail fast if unavailable
+PRESET="asr"
+JOBS=1
+FORCE=false
+QUIET=false
+LOWPASS=""
+SUFFIX="_clean"
+HIGHPASS="80"
+AFFTDN_NF="-25"      # noise floor in dB for afftdn
+AFFTDN_MD="8"        # mode for afftdn (higher can be more aggressive); requires builds that support 'md'
+NO_ADVANCED=false     # when true, avoid advanced options that some ffmpeg builds lack
+
+# Parse args
+if [[ $# -lt 1 ]]; then
+  print_usage
+  exit 1
+fi
+
+INPUT_PATH="$1"; shift || true
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -O|--out-dir)
+      OUT_DIR="$2"; shift 2;;
+    -e|--ext)
+      OUT_EXT="$2"; shift 2;;
+    -m|--model)
+      RN_MODEL="$2"; shift 2;;
+    --no-ml)
+      NO_ML=true; shift;;
+    --preset)
+      PRESET="$2"; shift 2;;
+    -j|--jobs)
+      JOBS="$2"; shift 2;;
+    -f|--force)
+      FORCE=true; shift;;
+    -q|--quiet)
+      QUIET=true; shift;;
+    --lowpass)
+      LOWPASS="$2"; shift 2;;
+    --suffix)
+      SUFFIX="$2"; shift 2;;
+    --no-advanced|--compat)
+      NO_ADVANCED=true; shift;;
+    --allow-fallback)
+      REQUIRE_ML=false; shift;;
+    -h|--help)
+      print_usage; exit 0;;
+    *)
+      echo "Unknown option: $1" >&2
+      print_usage
+      exit 1;;
+  esac
+done
+
+require_cmd ffmpeg
+
+# Resolve FFmpeg binary (env override -> local build -> system)
+FFMPEG_BIN=${FFMPEG_BIN:-}
+if [[ -z "${FFMPEG_BIN}" ]]; then
+  if [[ -x "$SCRIPT_DIR/ffmpeg-build/FFmpeg/ffmpeg" ]]; then
+    FFMPEG_BIN="$SCRIPT_DIR/ffmpeg-build/FFmpeg/ffmpeg"
+  else
+    FFMPEG_BIN="ffmpeg"
+  fi
+fi
+
+if ! command -v "$FFMPEG_BIN" >/dev/null 2>&1 && [[ ! -x "$FFMPEG_BIN" ]]; then
+  echo "Error: FFmpeg binary not found: $FFMPEG_BIN" >&2
+  exit 1
+fi
+if ! $QUIET; then
+  echo "Using FFmpeg binary: $FFMPEG_BIN" >&2
+fi
+
+FFMPEG_LOG=(-hide_banner)
+if $QUIET; then
+  FFMPEG_LOG+=( -loglevel error )
+else
+  FFMPEG_LOG+=( -loglevel info )
+fi
+
+FFMPEG_OVERWRITE=(-n)
+if $FORCE; then
+  FFMPEG_OVERWRITE=(-y)
+fi
+
+arnndn_available=false
+if "$FFMPEG_BIN" -hide_banner -h filter=arnndn >/dev/null 2>&1; then
+  arnndn_available=true
+else
+  if "$FFMPEG_BIN" -hide_banner -filters 2>/dev/null | grep -Eq '(^|[[:space:]])arnndn([[:space:]]|$)'; then
+    arnndn_available=true
+  fi
+fi
+if ! $QUIET; then
+  echo "arnndn_available=$arnndn_available" >&2
+fi
+
+# Check if afftdn supports 'md' option
+afftdn_supports_md=false
+if "$FFMPEG_BIN" -hide_banner -h filter=afftdn 2>/dev/null | grep -q " md="; then
+  afftdn_supports_md=true
+fi
+
+# Try to auto-discover an RNNoise model if none provided
+find_default_rn_model() {
+  local candidate=""
+  # Allow env variable override
+  if [[ -n "${RNNOISE_MODEL:-}" && -f "${RNNOISE_MODEL}" ]]; then
+    echo "${RNNOISE_MODEL}"
+    return 0
+  fi
+  local dirs=(
+    "$SCRIPT_DIR/models"
+    "$SCRIPT_DIR/../models"
+    "/usr/share/rnnoise"
+    "/usr/local/share/rnnoise"
+    "/usr/share/ffmpeg/models"
+    "$HOME/.local/share/rnnoise"
+  )
+  # Prefer '.rnnn' models (rnnoise-nu style) over legacy '.nn'
+  local exts=("rnnn" "nn" "model")
+  for d in "${dirs[@]}"; do
+    if [[ -d "$d" ]]; then
+      for ext in "${exts[@]}"; do
+        # Pick the first matching model file
+        for f in "$d"/*."$ext"; do
+          if [[ -f "$f" ]]; then
+            echo "$f"
+            return 0
+          fi
+        done
+      done
+    fi
+  done
+  return 1
+}
+
+use_arnndn=false
+if [[ $NO_ML == false ]]; then
+  if [[ $arnndn_available == false ]]; then
+    if $REQUIRE_ML; then
+      echo "Error: FFmpeg 'arnndn' filter not available. Please install/upgrade FFmpeg with librnnoise (see Bash/install_ffmpeg_with_arnndn.sh)." >&2
+      exit 9
+    fi
+  else
+    # arnndn available; require an external model
+    if [[ -n "$RN_MODEL" && -f "$RN_MODEL" ]]; then
+      :
+    else
+      if model_path=$(find_default_rn_model); then
+        RN_MODEL="$model_path"
+      else
+        if [[ -x "$SCRIPT_DIR/get_rnnoise_model.sh" ]]; then
+          RN_TARGET_DIR="$SCRIPT_DIR/models" RN_TARGET_NAME="rnnoise_model.rnnn" "$SCRIPT_DIR/get_rnnoise_model.sh" --yes || true
+          if model_path=$(find_default_rn_model); then
+            RN_MODEL="$model_path"
+          fi
+        fi
+      fi
+    fi
+    if [[ -z "$RN_MODEL" ]]; then
+      echo "Error: RNNoise model required but not found. Automatic download failed." >&2
+      echo "Hint: Set RN_URL to a reachable model URL and run Bash/get_rnnoise_model.sh, or supply -m /path/to/model.nn." >&2
+      exit 10
+    fi
+    use_arnndn=true
+    echo "Using RNNoise external model: $RN_MODEL" >&2
+  fi
+fi
+
+build_filters() {
+  local filters=()
+  # Remove low-frequency rumble typical for handheld/room noise
+  filters+=("highpass=f=${HIGHPASS}")
+
+  # Denoise
+  if $use_arnndn; then
+    # arnndn with full mix keeps the model output; if no external model, rely on built-in
+    filters+=("aresample=48000")
+    filters+=("arnndn=m=${RN_MODEL}:mix=1.0")
+  else
+    # afftdn: FFT-based denoise, tune nf (noise floor) as needed
+    if $REQUIRE_ML; then
+      echo "Error: RNNoise required but not in use; aborting rather than falling back to afftdn. Use --allow-fallback to permit." >&2
+      exit 11
+    fi
+    if $NO_ADVANCED; then
+      filters+=("afftdn=nf=${AFFTDN_NF}")
+    else
+      if $afftdn_supports_md; then
+        filters+=("afftdn=nf=${AFFTDN_NF}:md=${AFFTDN_MD}")
+      else
+        echo "Error: Your ffmpeg's afftdn filter does not support 'md='." >&2
+        echo "Hint: Install/upgrade ffmpeg to a build that supports afftdn md or rerun with --no-advanced." >&2
+        echo "      On Debian/Ubuntu you may need a newer ffmpeg from a PPA or build from source." >&2
+        exit 8
+      fi
+    fi
+  fi
+
+  # Optional low-pass to shave hiss; keep disabled unless requested
+  if [[ -n "$LOWPASS" ]]; then
+    filters+=("lowpass=f=${LOWPASS}")
+  fi
+
+  case "$PRESET" in
+    asr)
+      # ASR-friendly: avoid heavy gating/leveling, just prevent clipping
+      filters+=("alimiter=limit=0.94")
+      ;;
+    podcast)
+      # Gentle dynamic normalization and broadcast-ish loudness (single-pass)
+      # Note: single-pass loudnorm is approximate but OK for quick workflows
+      filters+=("dynaudnorm=f=500:g=5:p=0.1")
+      filters+=("loudnorm=i=-18:lra=9:tp=-2.0")
+      ;;
+    aggressive)
+      # Heavier clean-up; may harm ASR slightly but suppress background more
+      filters+=("agate=threshold=0.012:ratio=2.5:release=200")
+      filters+=("dynaudnorm=f=400:g=7:p=0.1")
+      filters+=("loudnorm=i=-18:lra=9:tp=-2.0")
+      ;;
+    *) ;;
+  esac
+
+  # Resample and format at the end for ASR
+  filters+=("aresample=16000")
+  filters+=("aformat=channel_layouts=mono:sample_fmts=s16")
+
+  local IFS=","; echo "${filters[*]}"
+}
+
+make_out_path_for_file() {
+  local in_file="$1"
+  local base
+  base=$(basename -- "$in_file")
+  base="${base%.*}"
+  local out_base="${base}${SUFFIX}.${OUT_EXT}"
+  if [[ -n "$OUT_DIR" ]]; then
+    mkdir -p -- "$OUT_DIR"
+    echo "$OUT_DIR/$out_base"
+  else
+    local dir
+    dir=$(dirname -- "$in_file")
+    echo "$dir/$out_base"
+  fi
+}
+
+process_one() {
+  local in_file="$1"
+  local out_file
+  out_file=$(make_out_path_for_file "$in_file")
+
+  # Choose codec based on extension
+  local codec=( -c:a pcm_s16le )
+  if [[ "$OUT_EXT" == "flac" ]]; then
+    codec=( -c:a flac )
+  fi
+
+  local af
+  af=$(build_filters)
+
+  if [[ -f "$out_file" && $FORCE == false ]]; then
+    echo "Skip (exists): $out_file"
+    return 0
+  fi
+
+  echo "Cleaning: $in_file -> $out_file"
+  "$FFMPEG_BIN" "${FFMPEG_LOG[@]}" "${FFMPEG_OVERWRITE[@]}" -i "$in_file" -af "$af" "${codec[@]}" "$out_file"
+}
+
+# Concurrency helpers (bash >= 5 supports wait -n; fallback to sequential if not)
+supports_wait_n=false
+if [[ -n "${BASH_VERSINFO:-}" && ${BASH_VERSINFO[0]} -ge 5 ]]; then
+  supports_wait_n=true
+fi
+
+run_dir() {
+  local dir="$1"
+  # Common audio extensions (case-insensitive)
+  mapfile -d '' files < <(find "$dir" -type f \
+    \( -iname "*.wav" -o -iname "*.mp3" -o -iname "*.m4a" -o -iname "*.aac" -o -iname "*.flac" \
+       -o -iname "*.ogg" -o -iname "*.opus" -o -iname "*.wma" -o -iname "*.webm" \) -print0)
+
+  if [[ ${#files[@]} -eq 0 ]]; then
+    echo "No audio files found in: $dir"
+    return 0
+  fi
+
+  local running=0
+  for f in "${files[@]}"; do
+    if [[ "$JOBS" -le 1 || $supports_wait_n == false ]]; then
+      process_one "$f"
+    else
+      process_one "$f" &
+      ((running++))
+      if (( running >= JOBS )); then
+        wait -n || true
+        ((running--))
+      fi
+    fi
+  done
+
+  # Wait for any remaining background jobs
+  if (( JOBS > 1 )) && $supports_wait_n; then
+    wait || true
+  fi
+}
+
+main() {
+  # Sanity checks and notices
+  if [[ -n "$RN_MODEL" && $use_arnndn == false && $NO_ML == false ]]; then
+    echo "Note: arnndn filter not available in your ffmpeg or model missing — using afftdn." >&2
+  fi
+
+  if [[ -f "$INPUT_PATH" ]]; then
+    process_one "$INPUT_PATH"
+  elif [[ -d "$INPUT_PATH" ]]; then
+    run_dir "$INPUT_PATH"
+  else
+    echo "Error: Input path not found: $INPUT_PATH" >&2
+    exit 1
+  fi
+}
+
+main "$@"
--- a/get_rnnoise_model.sh
+++ b/get_rnnoise_model.sh
@ -0,0 +1,190 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# get_rnnoise_model.sh — fetch an RNNoise model into a local models dir
+#
+# Prefers known-good rnnoise-nu models. You can override with:
+#   RN_URL, RN_TARGET_DIR, RN_TARGET_NAME
+#
+# Usage:
+#   Bash/get_rnnoise_model.sh            # interactive download
+#   RN_TARGET_DIR=./models Bash/get_rnnoise_model.sh --yes
+
+ask_yes_no() {
+  read -r -p "$1 [y/N]: " ans || true
+  case "${ans:-}" in
+    y|Y|yes|YES) return 0;;
+    *) return 1;;
+  esac
+}
+
+has_cmd() { command -v "$1" >/dev/null 2>&1; }
+
+YES=false
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -y|--yes) YES=true; shift;;
+    *) echo "Unknown option: $1" >&2; exit 2;;
+  esac
+done
+
+RN_TARGET_DIR=${RN_TARGET_DIR:-"$(dirname "$0")/models"}
+RN_TARGET_NAME=${RN_TARGET_NAME:-"rnnoise_model.rnnn"}
+
+mkdir -p "$RN_TARGET_DIR"
+dest="$RN_TARGET_DIR/$RN_TARGET_NAME"
+
+if [[ -f "$dest" ]]; then
+  echo "Model already exists at: $dest"
+  exit 0
+fi
+
+if ! $YES; then
+  if ! ask_yes_no "Download RNNoise model to $dest?"; then
+    echo "Aborted."
+    exit 1
+  fi
+fi
+
+if ! has_cmd curl && ! has_cmd wget; then
+  echo "Error: Need curl or wget to download RNNoise model." >&2
+  exit 3
+fi
+
+# Priority 1: explicit URL
+if [[ -n "${RN_URL:-}" ]]; then
+  echo "Downloading RNNoise model from RN_URL: $RN_URL" >&2
+  tmp=$(mktemp)
+  if has_cmd curl; then
+    curl -fsSL "$RN_URL" -o "$tmp"
+  else
+    wget -qO "$tmp" "$RN_URL"
+  fi
+  if [[ -s "$tmp" ]]; then
+    mv "$tmp" "$dest"
+    echo "Saved RNNoise model to: $dest"
+    exit 0
+  fi
+  rm -f "$tmp" || true
+  echo "Warning: RN_URL download failed; continuing to fallback sources." >&2
+fi
+
+# Priority 2: rnnoise-nu known models (GregorR)
+NU_URLS=(
+  "https://raw.githubusercontent.com/GregorR/rnnoise-nu/master/src/models/sh.rnnn"
+  "https://raw.githubusercontent.com/GregorR/rnnoise-nu/master/src/models/lq.rnnn"
+  "https://raw.githubusercontent.com/GregorR/rnnoise-nu/master/src/models/mp.rnnn"
+  "https://raw.githubusercontent.com/GregorR/rnnoise-nu/master/src/models/bd.rnnn"
+  "https://raw.githubusercontent.com/GregorR/rnnoise-nu/master/src/models/cb.rnnn"
+)
+for u in "${NU_URLS[@]}"; do
+  echo "Attempting to download RNNoise model from: $u" >&2
+  tmp=$(mktemp)
+  if has_cmd curl; then
+    if curl -fsSL "$u" -o "$tmp"; then
+      if [[ -s "$tmp" ]]; then
+        mv "$tmp" "$dest"
+        echo "Saved RNNoise model to: $dest" >&2
+        exit 0
+      fi
+    fi
+  else
+    if wget -qO "$tmp" "$u"; then
+      if [[ -s "$tmp" ]]; then
+        mv "$tmp" "$dest"
+        echo "Saved RNNoise model to: $dest" >&2
+        exit 0
+      fi
+    fi
+  fi
+  rm -f "$tmp" || true
+done
+
+# Priority 2b: arnndn-models fallback (richardpl)
+RNNDN_URLS=(
+  "https://raw.githubusercontent.com/richardpl/arnndn-models/master/sh.rnnn"
+)
+for u in "${RNNDN_URLS[@]}"; do
+  echo "Attempting to download RNNoise model from: $u" >&2
+  tmp=$(mktemp)
+  if has_cmd curl; then
+    if curl -fsSL "$u" -o "$tmp"; then
+      if [[ -s "$tmp" ]]; then
+        mv "$tmp" "$dest"
+        echo "Saved RNNoise model to: $dest" >&2
+        exit 0
+      fi
+    fi
+  else
+    if wget -qO "$tmp" "$u"; then
+      if [[ -s "$tmp" ]]; then
+        mv "$tmp" "$dest"
+        echo "Saved RNNoise model to: $dest" >&2
+        exit 0
+      fi
+    fi
+  fi
+  rm -f "$tmp" || true
+done
+
+# Priority 3: repo archives (rnnoise-nu and arnndn-models)
+ARCHIVES=(
+  "https://github.com/GregorR/rnnoise-nu/archive/refs/heads/master.zip"
+  "https://github.com/richardpl/arnndn-models/archive/refs/heads/master.zip"
+)
+for aurl in "${ARCHIVES[@]}"; do
+  echo "Attempting to download archive: $aurl" >&2
+  tmpdir=$(mktemp -d)
+  archive="$tmpdir/models.zip"
+  set +e
+  if has_cmd curl; then
+    curl -fL "$aurl" -o "$archive"
+  else
+    wget -O "$archive" "$aurl"
+  fi
+  status=$?
+  set -e
+  if [[ $status -ne 0 ]]; then
+    rm -rf "$tmpdir" || true
+    continue
+  fi
+  if has_cmd bsdtar; then
+    bsdtar -xf "$archive" -C "$tmpdir"
+  elif has_cmd unzip; then
+    unzip -q "$archive" -d "$tmpdir"
+  else
+    echo "Warning: Need bsdtar or unzip to extract archive; skipping archive method." >&2
+    rm -rf "$tmpdir" || true
+    continue
+  fi
+  mapfile -t nnfiles < <(bash -lc 'shopt -s globstar nullglob; for f in '"$tmpdir"'/**/*.rnnn '"$tmpdir"'/**/*.nn; do [[ -f "$f" ]] && echo "$f"; done')
+  if [[ ${#nnfiles[@]} -gt 0 ]]; then
+    cp -f "${nnfiles[0]}" "$dest"
+    echo "Saved RNNoise model to: $dest (from archive)" >&2
+    rm -rf "$tmpdir" || true
+    exit 0
+  fi
+  rm -rf "$tmpdir" || true
+done
+
+# Priority 4: Arch-based AUR packages and search only .nn/.rnnn
+if has_cmd yay; then
+  echo "Attempting to install AUR packages that may include RNNoise models..." >&2
+  set +e
+  yay -S --noconfirm denoiseit-git 2>/dev/null
+  yay -S --noconfirm speech-denoiser-git 2>/dev/null
+  set -e
+  mapfile -t found < <(bash -lc 'shopt -s globstar nullglob; for f in /usr/share/**/*.nn /usr/share/**/*.rnnn /usr/local/share/**/*.nn /usr/local/share/**/*.rnnn; do [[ -f "$f" ]] && echo "$f"; done' 2>/dev/null || true)
+  if [[ ${#found[@]} -gt 0 ]]; then
+    echo "Found candidate models:" >&2
+    printf '  %s\n' "${found[@]}" >&2
+    cp -f "${found[0]}" "$dest"
+    echo "Copied model to: $dest" >&2
+    exit 0
+  fi
+fi
+
+echo "Error: Could not obtain an RNNoise model automatically." >&2
+echo "Hint: Set RN_URL to a reachable model URL, or place a model file at: $dest" >&2
+exit 5
--- a/install_ffmpeg_with_arnndn.sh
+++ b/install_ffmpeg_with_arnndn.sh
@ -0,0 +1,130 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# install_ffmpeg_with_arnndn.sh — helper to install/upgrade FFmpeg with arnndn and full audio filters
+#
+# Tries distro packages first; if not suitable, offers to build from source.
+# This script prints commands and asks for confirmation before building.
+
+print_info() {
+  echo "[info] $*"
+}
+
+ask_yes_no() {
+  read -r -p "$1 [y/N]: " ans || true
+  case "${ans:-}" in
+    y|Y|yes|YES) return 0;;
+    *) return 1;;
+  esac
+}
+
+has_cmd() { command -v "$1" >/dev/null 2>&1; }
+
+detect_distro() {
+  if [[ -f /etc/os-release ]]; then
+    . /etc/os-release
+    echo "${ID:-unknown}"
+  else
+    echo "unknown"
+  fi
+}
+
+main() {
+  local distro
+  distro=$(detect_distro)
+  print_info "Detected distro: $distro"
+
+  if has_cmd ffmpeg && ffmpeg -hide_banner -filters | grep -q " arnndn "; then
+    print_info "Your ffmpeg already supports arnndn."
+  else
+    case "$distro" in
+      ubuntu|debian)
+        print_info "On Ubuntu/Debian, the official repo may lack newer filters. Consider a PPA or build from source."
+        echo "Options:"
+        echo "  - ppa: sudo add-apt-repository ppa:savoury1/ffmpeg6 && sudo apt update && sudo apt install ffmpeg"
+        echo "  - source build (recommended for latest): run this script to build from source"
+        ;;
+      arch|manjaro|endeavouros)
+        print_info "On Arch-based distros, ffmpeg is recent. Try: sudo pacman -Syu ffmpeg"
+        ;;
+      fedora)
+        print_info "On Fedora, try: sudo dnf install ffmpeg"
+        ;;
+      *)
+        print_info "Distro not recognized; will offer source build."
+        ;;
+    esac
+  fi
+
+  if ask_yes_no "Build FFmpeg from source with rnnoise/arnndn support now?"; then
+    echo "This will clone FFmpeg and build locally under ./ffmpeg-build. Continue?"
+    if ! ask_yes_no "Proceed"; then
+      exit 0
+    fi
+    set -x
+    mkdir -p ffmpeg-build && cd ffmpeg-build
+    # Prepare repository
+    if [[ -d FFmpeg ]]; then
+      if [[ -d FFmpeg/.git ]]; then
+        if ask_yes_no "An existing FFmpeg source directory was found. Reuse and update it?"; then
+          set +e
+          git -C FFmpeg fetch --all --tags --prune
+          git -C FFmpeg pull --rebase --ff-only || true
+          set -e
+        else
+          if ask_yes_no "Delete existing FFmpeg directory and re-clone?"; then
+            rm -rf FFmpeg
+          else
+            echo "Keeping existing FFmpeg directory as-is."
+          fi
+        fi
+      else
+        if ask_yes_no "Non-git 'FFmpeg' directory exists. Delete and re-clone?"; then
+          rm -rf FFmpeg
+        else
+          echo "Cannot proceed with a non-git FFmpeg directory present. Aborting."
+          exit 4
+        fi
+      fi
+    fi
+    # Dependencies
+    if [[ "$distro" == "ubuntu" || "$distro" == "debian" ]]; then
+      sudo apt update
+      sudo apt install -y git build-essential yasm nasm pkg-config libx264-dev libx265-dev libvpx-dev libopus-dev libfdk-aac-dev libmp3lame-dev libvorbis-dev libass-dev libfreetype6-dev libgnutls28-dev libaom-dev libdav1d-dev libxvidcore-dev libxcb1-dev libxcb-shm0-dev libxcb-xfixes0-dev libxcb-shape0-dev libdrm-dev libvulkan-dev libva-dev libvdpau-dev librtmp-dev libunistring-dev libgnutls28-dev libchromaprint-dev libbluray-dev librubberband-dev libspeex-dev libsoxr-dev libvmaf-dev libzimg-dev libsvtav1-dev libtheora-dev libwebp-dev libopenal-dev libjack-jackd2-dev libpulse-dev librnnoise-dev
+    elif [[ "$distro" == "arch" || "$distro" == "manjaro" || "$distro" == "endeavouros" ]]; then
+      sudo pacman -Syu --needed base-devel yasm nasm pkgconf rnnoise
+    elif [[ "$distro" == "fedora" ]]; then
+      sudo dnf install -y git make gcc yasm nasm pkgconf-pkg-config rnnoise-devel libX11-devel libXext-devel libXfixes-devel libXv-devel libXrandr-devel libXi-devel libXtst-devel libXinerama-devel freetype-devel fontconfig-devel libass-devel libvpx-devel libaom-devel libdav1d-devel zimg-devel rubberband-devel soxr-devel libvorbis-devel opus-devel lame-devel
+    else
+      echo "Note: please ensure rnnoise development headers are installed (pkg-config rnnoise)." >&2
+    fi
+    if [[ ! -d FFmpeg/.git ]]; then
+      git clone https://github.com/FFmpeg/FFmpeg.git --depth=1
+    fi
+    cd FFmpeg
+    RN_FLAG=""
+    # Some FFmpeg versions auto-detect rnnoise without a flag; include the flag only if supported
+    if ./configure --help | grep -q "librnnoise"; then
+      RN_FLAG="--enable-librnnoise"
+    else
+      echo "[info] configure has no --enable-librnnoise; relying on auto-detection via pkg-config (rnnoise)." >&2
+    fi
+
+    ./configure \
+      --enable-gpl --enable-nonfree \
+      --enable-libx264 --enable-libx265 --enable-libvpx --enable-libopus --enable-libmp3lame \
+      --enable-libvorbis --enable-libass --enable-fontconfig --enable-libfreetype \
+      --enable-librubberband --enable-libsoxr --enable-libzimg --enable-libvmaf \
+      --enable-libdav1d --enable-libaom --enable-libsvtav1 \
+      ${RN_FLAG} \
+      --enable-ffplay --enable-ffprobe
+    make -j"$(nproc)"
+    echo "Build complete. You can run ./ffmpeg-build/FFmpeg/ffmpeg from this folder or 'sudo make install' to install system-wide."
+    set +x
+  else
+    echo "Skipped building from source."
+  fi
+}
+
+main "$@"