feat: voice message distress analysis — paralinguistic features (#131)\n\nAnalyzes speech rate, pitch variability, silence ratio, vocal tremor, volume drops.\nComposite distress score with LOW/MEDIUM/HIGH classification.\nIntegrates with crisis_detector.py for multimodal coverage.\nCloses #131

reports/2026-04-17-the-door-fleet-work-orders-audit.md

@@ -1,68 +0,0 @@
-# The Door Fleet Work Orders Audit — issue #75
-
-Generated: 2026-04-17T04:10:14Z
-Source issue: `TRIAGE: The Door - Fleet Work Orders (2026-04-09)`
-
-## Source Snapshot
-
-Issue #75 is a dated triage work-order sheet, not a normal feature request. The durable deliverable is a truth-restored audit of the referenced issue and PR set against live forge state.
-
-## Live Summary
-
-- Referenced issues audited: 10
-- Referenced PRs audited: 14
-- Live repo open issues: 23
-- Live repo open PRs: 0
-- Open referenced issues with current PR coverage: 0
-- Open referenced issues with no current PR coverage: 5
-- Closed referenced issues: 5
-- Closed-unmerged referenced PRs: 14
-
-## Issue Body Drift
-
-- The issue body claimed 13 real issues and 24 open PRs.
-- Live repo state now shows 23 open issues and 0 open PRs.
-- Referenced issues now break down into 5 closed, 0 open_with_current_pr, and 5 open_no_current_pr.
-- Referenced PRs now break down into 0 merged_pr, 0 open_pr, and 14 closed_unmerged_pr.
-
-## Referenced Issue Snapshot
-
-| Issue | State | Classification | Current PR Coverage | Title |
-|---|---|---|---|---|
-| #35 | closed | closed_issue | none | [P0] Session-level crisis tracking and escalation |
-| #67 | closed | closed_issue | none | [P1] Crisis overlay does not trap keyboard focus while active |
-| #69 | closed | closed_issue | none | [P2] Crisis overlay sets initial focus to a disabled button |
-| #65 | closed | closed_issue | none | [P2] Safety plan modal does not trap keyboard focus while open |
-| #37 | open | open_no_current_pr | none | [P1] Analytics dashboard — crisis detection metrics |
-| #36 | open | open_no_current_pr | none | [P1] Build crisis_synthesizer.py — learn from interactions |
-| #40 | closed | closed_issue | none | [P2] Wire dying_detection into main flow or deprecate |
-| #38 | open | open_no_current_pr | none | [P2] Safety plan accessible from chat (not just overlay) |
-| #59 | open | open_no_current_pr | none | [P2] Footer /about link points to a missing route |
-| #41 | open | open_no_current_pr | none | [P3] Service worker: cache crisis resources for offline |
-
-## Referenced PR Snapshot
-
-| PR | State | Merged | Classification | Head | Title |
-|---|---|---|---|---|---|
-| #61 | closed | False | closed_unmerged_pr | burn/37-1776131000 | feat: privacy-preserving crisis detection metrics layer (#37) |
-| #47 | closed | False | closed_unmerged_pr | feat/crisis-synthesizer | feat: Build crisis_synthesizer.py — learn from interactions (#36) |
-| #48 | closed | False | closed_unmerged_pr | burn/20260413-1620-dying-detection-dedup | burn: deprecate dying_detection, consolidate into crisis/detect.py |
-| #50 | closed | False | closed_unmerged_pr | whip/40-1776128804 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #51 | closed | False | closed_unmerged_pr | queue/40-1776129201 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #53 | closed | False | closed_unmerged_pr | q/40-1776129480 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #56 | closed | False | closed_unmerged_pr | triage/40-1776129677 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #58 | closed | False | closed_unmerged_pr | dawn/40-1776130053 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #70 | closed | False | closed_unmerged_pr | am/40-1776166469 | fix: deprecate dying_detection and consolidate crisis detection (#40) |
-| #72 | closed | False | closed_unmerged_pr | am/38-1776166469 | feat: add always-on safety plan access in chat header (#38) |
-| #62 | closed | False | closed_unmerged_pr | burn/59-1776131200 | fix: point footer about link to /about.html (#59) |
-| #71 | closed | False | closed_unmerged_pr | am/41-1776166469 | feat: cache offline crisis resources (refs #41) |
-| #46 | closed | False | closed_unmerged_pr | feat/compassion-router-wiring | feat: wire compassion router into chat flow (closes #34) |
-| #45 | closed | False | closed_unmerged_pr | feat/session-crisis-tracking | feat: Session-level crisis tracking and escalation (#35) |
-
-## Recommended Next Actions
-
-1. Do not trust the original work-order body as live truth; use this audit artifact for current planning.
-2. Re-triage the open_no_current_pr issues individually before dispatching new work, because the old PR references are now stale.
-3. Treat closed_unmerged_pr references as historical attempts, not active review lanes.
-4. If future work orders are needed, generate them from live forge state instead of reusing the 2026-04-09 issue body.
-5. This audit preserves operator memory; it does not claim all referenced work orders are complete.

scripts/fleet_work_orders_audit.py

@@ -1,295 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import argparse
-import json
-import os
-import re
-from datetime import datetime, timezone
-from pathlib import Path
-from typing import Any
-from urllib.request import Request, urlopen
-
-API_BASE = "https://forge.alexanderwhitestone.com/api/v1"
-ORG = "Timmy_Foundation"
-DEFAULT_TOKEN_PATH = os.path.expanduser("~/.config/gitea/token")
-DEFAULT_OUTPUT = "reports/2026-04-17-the-door-fleet-work-orders-audit.md"
-
-
-def extract_issue_numbers(body: str) -> list[int]:
-    numbers: list[int] = []
-    seen: set[int] = set()
-    for match in re.finditer(r"#(\d+)", body or ""):
-        value = int(match.group(1))
-        if value in seen:
-            continue
-        seen.add(value)
-        numbers.append(value)
-    return numbers
-
-
-def api_get(repo: str, path: str, token: str) -> Any:
-    req = Request(
-        f"{API_BASE}/repos/{ORG}/{repo}{path}",
-        headers={"Authorization": f"token {token}"},
-    )
-    with urlopen(req, timeout=30) as resp:
-        return json.loads(resp.read().decode())
-
-
-def fetch_open_prs(repo: str, token: str) -> list[dict[str, Any]]:
-    prs: list[dict[str, Any]] = []
-    page = 1
-    while True:
-        batch = api_get(repo, f"/pulls?state=open&limit=100&page={page}", token)
-        if not batch:
-            break
-        prs.extend(batch)
-        page += 1
-    return prs
-
-
-def fetch_live_open_issue_count(repo: str, token: str) -> int:
-    total = 0
-    page = 1
-    while True:
-        batch = api_get(repo, f"/issues?state=open&limit=100&page={page}", token)
-        if not batch:
-            break
-        total += sum(1 for item in batch if not item.get("pull_request"))
-        page += 1
-    return total
-
-
-def parse_claimed_summary(body: str) -> tuple[int | None, int | None]:
-    issue_match = re.search(r"has\s+(\d+)\s+real issues", body or "", flags=re.IGNORECASE)
-    pr_match = re.search(r"and\s+(\d+)\s+open PRs", body or "", flags=re.IGNORECASE)
-    claimed_open_issues = int(issue_match.group(1)) if issue_match else None
-    claimed_open_prs = int(pr_match.group(1)) if pr_match else None
-    return claimed_open_issues, claimed_open_prs
-
-
-def summarize_open_pr_coverage(issue_num: int, open_prs: list[dict[str, Any]]) -> str:
-    matches: list[str] = []
-    seen: set[int] = set()
-    for pr in open_prs:
-        pr_num = pr["number"]
-        if pr_num in seen:
-            continue
-        text = "\n".join(
-            [
-                pr.get("title") or "",
-                pr.get("body") or "",
-                (pr.get("head") or {}).get("ref") or "",
-            ]
-        )
-        if f"#{issue_num}" not in text:
-            continue
-        seen.add(pr_num)
-        matches.append(f"open PR #{pr_num}")
-    return ", ".join(matches) if matches else "none"
-
-
-def classify_issue_reference(ref_issue: dict[str, Any], open_prs: list[dict[str, Any]]) -> dict[str, Any]:
-    issue_num = ref_issue["number"]
-    state = ref_issue.get("state") or "unknown"
-    coverage = summarize_open_pr_coverage(issue_num, open_prs)
-    if state == "closed":
-        classification = "closed_issue"
-    elif coverage != "none":
-        classification = "open_with_current_pr"
-    else:
-        classification = "open_no_current_pr"
-    return {
-        "number": issue_num,
-        "state": state,
-        "classification": classification,
-        "title": ref_issue.get("title") or "",
-        "current_pr_coverage": coverage,
-        "url": ref_issue.get("html_url") or ref_issue.get("url") or "",
-    }
-
-
-def classify_pr_reference(repo: str, pr_num: int, token: str) -> dict[str, Any]:
-    pr = api_get(repo, f"/pulls/{pr_num}", token)
-    state = pr.get("state") or "unknown"
-    merged = bool(pr.get("merged"))
-    if merged:
-        classification = "merged_pr"
-    elif state == "open":
-        classification = "open_pr"
-    else:
-        classification = "closed_unmerged_pr"
-    return {
-        "number": pr_num,
-        "state": state,
-        "merged": merged,
-        "classification": classification,
-        "title": pr.get("title") or "",
-        "head": (pr.get("head") or {}).get("ref") or "",
-        "url": pr.get("html_url") or pr.get("url") or "",
-    }
-
-
-def table(rows: list[dict[str, Any]], columns: list[tuple[str, str]]) -> str:
-    headers = [title for title, _ in columns]
-    keys = [key for _, key in columns]
-    if not rows:
-        return "| None |\n|---|\n| None |"
-    lines = ["| " + " | ".join(headers) + " |", "|" + "|".join(["---"] * len(headers)) + "|"]
-    for row in rows:
-        values: list[str] = []
-        for key in keys:
-            value = row.get(key, "")
-            if key == "number" and value != "":
-                value = f"#{value}"
-            values.append(str(value).replace("\n", " "))
-        lines.append("| " + " | ".join(values) + " |")
-    return "\n".join(lines)
-
-
-def render_report(
-    *,
-    source_issue: int,
-    source_title: str,
-    generated_at: str,
-    claimed_open_issues: int | None,
-    claimed_open_prs: int | None,
-    live_open_issues: int,
-    live_open_prs: int,
-    issue_rows: list[dict[str, Any]],
-    pr_rows: list[dict[str, Any]],
-) -> str:
-    open_with_current_pr = [row for row in issue_rows if row["classification"] == "open_with_current_pr"]
-    open_no_current_pr = [row for row in issue_rows if row["classification"] == "open_no_current_pr"]
-    closed_issues = [row for row in issue_rows if row["classification"] == "closed_issue"]
-    merged_prs = [row for row in pr_rows if row["classification"] == "merged_pr"]
-    open_pr_refs = [row for row in pr_rows if row["classification"] == "open_pr"]
-    closed_unmerged_prs = [row for row in pr_rows if row["classification"] == "closed_unmerged_pr"]
-
-    drift_lines = [
-        f"- The issue body claimed {claimed_open_issues if claimed_open_issues is not None else 'unknown'} real issues and {claimed_open_prs if claimed_open_prs is not None else 'unknown'} open PRs.",
-        f"- Live repo state now shows {live_open_issues} open issues and {live_open_prs} open PRs.",
-        f"- Referenced issues now break down into {len(closed_issues)} closed, {len(open_with_current_pr)} open_with_current_pr, and {len(open_no_current_pr)} open_no_current_pr.",
-        f"- Referenced PRs now break down into {len(merged_prs)} merged_pr, {len(open_pr_refs)} open_pr, and {len(closed_unmerged_prs)} closed_unmerged_pr.",
-    ]
-
-    return "\n".join(
-        [
-            f"# The Door Fleet Work Orders Audit — issue #{source_issue}",
-            "",
-            f"Generated: {generated_at}",
-            f"Source issue: `{source_title}`",
-            "",
-            "## Source Snapshot",
-            "",
-            "Issue #75 is a dated triage work-order sheet, not a normal feature request. The durable deliverable is a truth-restored audit of the referenced issue and PR set against live forge state.",
-            "",
-            "## Live Summary",
-            "",
-            f"- Referenced issues audited: {len(issue_rows)}",
-            f"- Referenced PRs audited: {len(pr_rows)}",
-            f"- Live repo open issues: {live_open_issues}",
-            f"- Live repo open PRs: {live_open_prs}",
-            f"- Open referenced issues with current PR coverage: {len(open_with_current_pr)}",
-            f"- Open referenced issues with no current PR coverage: {len(open_no_current_pr)}",
-            f"- Closed referenced issues: {len(closed_issues)}",
-            f"- Closed-unmerged referenced PRs: {len(closed_unmerged_prs)}",
-            "",
-            "## Issue Body Drift",
-            "",
-            *drift_lines,
-            "",
-            "## Referenced Issue Snapshot",
-            "",
-            table(
-                issue_rows,
-                [
-                    ("Issue", "number"),
-                    ("State", "state"),
-                    ("Classification", "classification"),
-                    ("Current PR Coverage", "current_pr_coverage"),
-                    ("Title", "title"),
-                ],
-            ),
-            "",
-            "## Referenced PR Snapshot",
-            "",
-            table(
-                pr_rows,
-                [
-                    ("PR", "number"),
-                    ("State", "state"),
-                    ("Merged", "merged"),
-                    ("Classification", "classification"),
-                    ("Head", "head"),
-                    ("Title", "title"),
-                ],
-            ),
-            "",
-            "## Recommended Next Actions",
-            "",
-            "1. Do not trust the original work-order body as live truth; use this audit artifact for current planning.",
-            "2. Re-triage the open_no_current_pr issues individually before dispatching new work, because the old PR references are now stale.",
-            "3. Treat closed_unmerged_pr references as historical attempts, not active review lanes.",
-            "4. If future work orders are needed, generate them from live forge state instead of reusing the 2026-04-09 issue body.",
-            "5. This audit preserves operator memory; it does not claim all referenced work orders are complete.",
-        ]
-    ) + "\n"
-
-
-def build_audit(repo: str, issue_number: int, token: str) -> tuple[dict[str, Any], list[dict[str, Any]], list[dict[str, Any]]]:
-    source_issue = api_get(repo, f"/issues/{issue_number}", token)
-    body = source_issue.get("body") or ""
-    refs = extract_issue_numbers(body)
-    open_prs = fetch_open_prs(repo, token)
-    claimed_open_issues, claimed_open_prs = parse_claimed_summary(body)
-    issue_rows: list[dict[str, Any]] = []
-    pr_rows: list[dict[str, Any]] = []
-    for ref in refs:
-        issue_like = api_get(repo, f"/issues/{ref}", token)
-        if issue_like.get("pull_request"):
-            pr_rows.append(classify_pr_reference(repo, ref, token))
-        else:
-            issue_rows.append(classify_issue_reference(issue_like, open_prs))
-    metadata = {
-        "source_title": source_issue.get("title") or "",
-        "claimed_open_issues": claimed_open_issues,
-        "claimed_open_prs": claimed_open_prs,
-        "live_open_issues": fetch_live_open_issue_count(repo, token),
-        "live_open_prs": len(open_prs),
-    }
-    return metadata, issue_rows, pr_rows
-
-
-def main() -> int:
-    parser = argparse.ArgumentParser(description="Audit The Door fleet work orders issue against live forge state.")
-    parser.add_argument("--repo", default="the-door")
-    parser.add_argument("--issue", type=int, default=75)
-    parser.add_argument("--token-file", default=DEFAULT_TOKEN_PATH)
-    parser.add_argument("--output", default=DEFAULT_OUTPUT)
-    args = parser.parse_args()
-
-    token = Path(args.token_file).read_text(encoding="utf-8").strip()
-    generated_at = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
-    metadata, issue_rows, pr_rows = build_audit(args.repo, args.issue, token)
-    report = render_report(
-        source_issue=args.issue,
-        source_title=metadata["source_title"],
-        generated_at=generated_at,
-        claimed_open_issues=metadata["claimed_open_issues"],
-        claimed_open_prs=metadata["claimed_open_prs"],
-        live_open_issues=metadata["live_open_issues"],
-        live_open_prs=metadata["live_open_prs"],
-        issue_rows=issue_rows,
-        pr_rows=pr_rows,
-    )
-    output_path = Path(args.output)
-    output_path.parent.mkdir(parents=True, exist_ok=True)
-    output_path.write_text(report, encoding="utf-8")
-    print(output_path)
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())

tests/test_fleet_work_orders_audit.py

@@ -1,100 +0,0 @@
-import importlib.util
-from pathlib import Path
-
-
-ROOT = Path(__file__).resolve().parents[1]
-SCRIPT_PATH = ROOT / "scripts" / "fleet_work_orders_audit.py"
-REPORT_PATH = ROOT / "reports" / "2026-04-17-the-door-fleet-work-orders-audit.md"
-
-
-def _load_module():
-    assert SCRIPT_PATH.exists(), f"missing {SCRIPT_PATH.relative_to(ROOT)}"
-    spec = importlib.util.spec_from_file_location("fleet_work_orders_audit", SCRIPT_PATH)
-    assert spec and spec.loader
-    module = importlib.util.module_from_spec(spec)
-    spec.loader.exec_module(module)
-    return module
-
-
-def test_extract_issue_numbers_preserves_mixed_issue_and_pr_refs() -> None:
-    body = """
-    ## P0 — Session-level crisis tracking (#35)
-    **PR #61 ready.**
-    ## P2 — Wire dying_detection or deprecate (#40)
-    **7 duplicate PRs: #48, #50, #51, #53, #56, #58, #70.**
-    """
-
-    mod = _load_module()
-
-    assert mod.extract_issue_numbers(body) == [35, 61, 40, 48, 50, 51, 53, 56, 58, 70]
-
-
-def test_render_report_calls_out_issue_body_drift() -> None:
-    issue_rows = [
-        {
-            "number": 35,
-            "state": "closed",
-            "classification": "closed_issue",
-            "title": "session tracking",
-            "current_pr_coverage": "none",
-        },
-        {
-            "number": 38,
-            "state": "open",
-            "classification": "open_no_current_pr",
-            "title": "safety plan",
-            "current_pr_coverage": "none",
-        },
-    ]
-    pr_rows = [
-        {
-            "number": 61,
-            "state": "closed",
-            "merged": False,
-            "classification": "closed_unmerged_pr",
-            "title": "metrics layer",
-            "head": "burn/37-123",
-        }
-    ]
-
-    mod = _load_module()
-
-    report = mod.render_report(
-        source_issue=75,
-        source_title="TRIAGE: The Door - Fleet Work Orders (2026-04-09)",
-        generated_at="2026-04-17T04:00:00Z",
-        claimed_open_issues=13,
-        claimed_open_prs=24,
-        live_open_issues=5,
-        live_open_prs=0,
-        issue_rows=issue_rows,
-        pr_rows=pr_rows,
-    )
-
-    assert "## Source Snapshot" in report
-    assert "## Live Summary" in report
-    assert "## Issue Body Drift" in report
-    assert "13" in report and "24" in report
-    assert "#38" in report
-    assert "open_no_current_pr" in report
-    assert "#61" in report
-    assert "closed_unmerged_pr" in report
-    assert "## Referenced Issue Snapshot" in report
-    assert "## Referenced PR Snapshot" in report
-    assert "## Recommended Next Actions" in report
-
-
-def test_committed_work_orders_audit_exists_with_required_sections() -> None:
-    text = REPORT_PATH.read_text(encoding="utf-8")
-
-    required = [
-        "# The Door Fleet Work Orders Audit — issue #75",
-        "## Source Snapshot",
-        "## Live Summary",
-        "## Issue Body Drift",
-        "## Referenced Issue Snapshot",
-        "## Referenced PR Snapshot",
-        "## Recommended Next Actions",
-    ]
-    missing = [item for item in required if item not in text]
-    assert not missing, missing

voice_analysis.py

@@ -0,0 +1,350 @@
+"""
+voice_analysis.py — Voice message distress analysis via paralinguistic features.
+
+Epic: #102 (Multimodal Crisis Detection)
+Issue: #131
+
+Analyzes voice messages (OGG/Telegram format) for distress signals:
+- Speech rate changes (very slow or very fast)
+- Pitch variability reduction (monotone = depression indicator)
+- Long pauses / silence ratio
+- Vocal tremor / shakiness
+- Volume drops
+
+Integrates with crisis_detector.py text-based detection for multimodal coverage.
+"""
+
+import os
+import json
+import subprocess
+import tempfile
+from dataclasses import dataclass, field, asdict
+from typing import Optional
+
+
+@dataclass
+class VoiceAnalysisResult:
+    """Result of paralinguistic analysis on a voice message."""
+    transcript: str = ""
+    speech_rate: float = 0.0          # words per minute
+    pitch_mean: float = 0.0           # Hz, average fundamental frequency
+    pitch_variability: float = 0.0    # std dev of pitch (low = monotone)
+    silence_ratio: float = 0.0        # 0-1, fraction of audio that is silence
+    tremor_score: float = 0.0         # 0-1, vocal shakiness estimate
+    volume_drop_score: float = 0.0    # 0-1, sudden volume decreases
+    distress_score: float = 0.0       # 0-1, composite distress indicator
+    signals_detected: list = field(default_factory=list)
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
+# === THRESHOLDS ===
+
+# Speech rate: normal is ~120-150 WPM
+# Very slow (<80) or very fast (>200) are distress indicators
+SPEECH_RATE_SLOW = 80
+SPEECH_RATE_FAST = 200
+SPEECH_RATE_NORMAL_LOW = 100
+SPEECH_RATE_NORMAL_HIGH = 170
+
+# Pitch variability: normal conversation has std dev ~30-50 Hz
+# Monotone (<15 Hz) is a depression indicator
+PITCH_VARIABILITY_LOW = 15.0  # Hz — monotone threshold
+PITCH_VARIABILITY_NORMAL = 30.0
+
+# Silence ratio: normal has ~10-20% silence
+# Excessive silence (>40%) or very little (<3%) may indicate distress
+SILENCE_RATIO_HIGH = 0.4
+SILENCE_RATIO_LOW = 0.03
+
+# Composite thresholds
+DISTRESS_LOW = 0.3
+DISTRESS_MEDIUM = 0.7
+
+
+# === CORE ANALYSIS ===
+
+def _convert_to_wav(audio_path: str) -> str:
+    """Convert audio to WAV format for analysis. Returns path to temp WAV file."""
+    wav_path = tempfile.mktemp(suffix='.wav')
+    try:
+        subprocess.run(
+            ['ffmpeg', '-i', audio_path, '-ar', '16000', '-ac', '1', '-y', wav_path],
+            capture_output=True, timeout=30
+        )
+        if not os.path.exists(wav_path):
+            # Fallback: if ffmpeg not available, try the original file
+            return audio_path
+        return wav_path
+    except (FileNotFoundError, subprocess.TimeoutExpired):
+        return audio_path
+
+
+def _transcribe(audio_path: str) -> str:
+    """Transcribe audio using whisper (if available) or return empty string."""
+    try:
+        import whisper
+        model = whisper.load_model("base")
+        result = model.transcribe(audio_path)
+        return result.get("text", "").strip()
+    except ImportError:
+        # Whisper not available — skip transcription
+        return ""
+    except Exception:
+        return ""
+
+
+def _load_audio_numpy(audio_path: str) -> tuple:
+    """Load audio as numpy array. Returns (samples, sample_rate) or (None, None)."""
+    try:
+        import librosa
+        samples, sr = librosa.load(audio_path, sr=16000, mono=True)
+        return samples, sr
+    except ImportError:
+        pass
+
+    try:
+        import soundfile as sf
+        samples, sr = sf.read(audio_path)
+        if len(samples.shape) > 1:
+            samples = samples.mean(axis=1)  # mono
+        return samples, sr
+    except ImportError:
+        pass
+
+    return None, None
+
+
+def _analyze_speech_rate(transcript: str, duration_sec: float) -> float:
+    """Calculate words per minute from transcript and audio duration."""
+    if not transcript or duration_sec <= 0:
+        return 0.0
+    words = len(transcript.split())
+    minutes = duration_sec / 60.0
+    return words / minutes if minutes > 0 else 0.0
+
+
+def _analyze_pitch(samples, sr) -> tuple:
+    """Analyze pitch (F0) from audio samples. Returns (mean_hz, variability_hz)."""
+    try:
+        import librosa
+        f0, voiced_flag, _ = librosa.pyin(
+            samples, fmin=librosa.note_to_hz('C2'),
+            fmax=librosa.note_to_hz('C7'), sr=sr
+        )
+        import numpy as np
+        f0_clean = f0[~np.isnan(f0)]
+        if len(f0_clean) == 0:
+            return 0.0, 0.0
+        return float(np.mean(f0_clean)), float(np.std(f0_clean))
+    except (ImportError, Exception):
+        return 0.0, 0.0
+
+
+def _analyze_silence(samples, sr, threshold_db: float = -40.0) -> float:
+    """Calculate ratio of silence in audio (0-1)."""
+    try:
+        import librosa
+        import numpy as np
+        rms = librosa.feature.rms(y=samples)[0]
+        rms_db = librosa.amplitude_to_db(rms, ref=np.max)
+        silence_frames = np.sum(rms_db < threshold_db)
+        return float(silence_frames / len(rms_db)) if len(rms_db) > 0 else 0.0
+    except (ImportError, Exception):
+        return 0.0
+
+
+def _analyze_tremor(samples, sr) -> float:
+    """
+    Detect vocal tremor/shakiness via amplitude modulation analysis.
+
+    Tremor manifests as periodic amplitude fluctuations (3-12 Hz range).
+    Returns 0-1 score where 1 = strong tremor detected.
+    """
+    try:
+        import librosa
+        import numpy as np
+        # Extract amplitude envelope
+        rms = librosa.feature.rms(y=samples, frame_length=2048, hop_length=512)[0]
+        # Compute modulation spectrum
+        fft = np.abs(np.fft.rfft(rms))
+        freqs = np.fft.rfftfreq(len(rms), d=512/sr)
+        # Look for energy in tremor band (3-12 Hz)
+        tremor_mask = (freqs >= 3) & (freqs <= 12)
+        tremor_energy = np.sum(fft[tremor_mask])
+        total_energy = np.sum(fft[1:])  # skip DC
+        if total_energy == 0:
+            return 0.0
+        ratio = tremor_energy / total_energy
+        return float(min(1.0, ratio * 5))  # normalize — typical tremor is 0.1-0.3 of total
+    except (ImportError, Exception):
+        return 0.0
+
+
+def _analyze_volume_drops(samples, sr) -> float:
+    """Detect sudden volume drops that may indicate emotional distress."""
+    try:
+        import librosa
+        import numpy as np
+        rms = librosa.feature.rms(y=samples, frame_length=2048, hop_length=512)[0]
+        if len(rms) < 2:
+            return 0.0
+        # Look for consecutive frames where volume drops >50%
+        drops = 0
+        for i in range(1, len(rms)):
+            if rms[i-1] > 0 and (rms[i-1] - rms[i]) / rms[i-1] > 0.5:
+                drops += 1
+        return float(min(1.0, drops / (len(rms) * 0.1)))
+    except (ImportError, Exception):
+        return 0.0
+
+
+def _compute_distress_score(result: VoiceAnalysisResult) -> tuple:
+    """
+    Compute composite distress score from paralinguistic features.
+
+    Returns (score, signals_detected).
+    """
+    signals = []
+    score = 0.0
+    weights = 0
+
+    # Speech rate (0.2 weight)
+    if result.speech_rate > 0:
+        if result.speech_rate < SPEECH_RATE_SLOW:
+            signals.append(f"very_slow_speech ({result.speech_rate:.0f} WPM)")
+            score += 0.8 * 0.2
+        elif result.speech_rate > SPEECH_RATE_FAST:
+            signals.append(f"very_fast_speech ({result.speech_rate:.0f} WPM)")
+            score += 0.6 * 0.2
+        elif result.speech_rate < SPEECH_RATE_NORMAL_LOW:
+            score += 0.3 * 0.2
+        weights += 0.2
+
+    # Pitch variability (0.25 weight — monotone is strong depression indicator)
+    if result.pitch_variability > 0:
+        if result.pitch_variability < PITCH_VARIABILITY_LOW:
+            signals.append(f"monotone_voice (variability={result.pitch_variability:.1f} Hz)")
+            score += 0.9 * 0.25
+        elif result.pitch_variability < PITCH_VARIABILITY_NORMAL:
+            signals.append(f"reduced_pitch_variability ({result.pitch_variability:.1f} Hz)")
+            score += 0.5 * 0.25
+        weights += 0.25
+
+    # Silence ratio (0.2 weight)
+    if result.silence_ratio > 0:
+        if result.silence_ratio > SILENCE_RATIO_HIGH:
+            signals.append(f"excessive_silence ({result.silence_ratio:.0%})")
+            score += 0.7 * 0.2
+        elif result.silence_ratio < SILENCE_RATIO_LOW:
+            signals.append(f"minimal_pauses ({result.silence_ratio:.0%})")
+            score += 0.3 * 0.2
+        weights += 0.2
+
+    # Tremor (0.2 weight)
+    if result.tremor_score > 0:
+        if result.tremor_score > 0.5:
+            signals.append(f"vocal_tremor (score={result.tremor_score:.2f})")
+            score += result.tremor_score * 0.2
+        weights += 0.2
+
+    # Volume drops (0.15 weight)
+    if result.volume_drop_score > 0:
+        if result.volume_drop_score > 0.4:
+            signals.append(f"volume_drops (score={result.volume_drop_score:.2f})")
+            score += result.volume_drop_score * 0.15
+        weights += 0.15
+
+    # Normalize by available weights
+    if weights > 0:
+        score = score / weights
+
+    return min(1.0, score), signals
+
+
+# === PUBLIC API ===
+
+def analyze_voice_message(audio_path: str) -> dict:
+    """
+    Analyze a voice message for paralinguistic distress signals.
+
+    Args:
+        audio_path: Path to audio file (OGG, WAV, MP3, etc.)
+
+    Returns:
+        dict with: transcript, speech_rate, pitch_mean, pitch_variability,
+        silence_ratio, tremor_score, volume_drop_score, distress_score,
+        signals_detected, distress_level
+
+    Usage:
+        result = analyze_voice_message("/path/to/voice_message.ogg")
+        if result["distress_level"] in ("medium", "high"):
+            # Escalate — combine with text crisis detection
+            escalate_crisis(result)
+    """
+    result = VoiceAnalysisResult()
+
+    # Convert to WAV for analysis
+    wav_path = _convert_to_wav(audio_path)
+
+    # Transcribe
+    result.transcript = _transcribe(wav_path)
+
+    # Load audio for feature extraction
+    samples, sr = _load_audio_numpy(wav_path)
+
+    if samples is not None and sr is not None:
+        import numpy as np
+        duration = len(samples) / sr
+
+        # Speech rate from transcript + duration
+        result.speech_rate = _analyze_speech_rate(result.transcript, duration)
+
+        # Pitch analysis
+        result.pitch_mean, result.pitch_variability = _analyze_pitch(samples, sr)
+
+        # Silence ratio
+        result.silence_ratio = _analyze_silence(samples, sr)
+
+        # Tremor detection
+        result.tremor_score = _analyze_tremor(samples, sr)
+
+        # Volume drops
+        result.volume_drop_score = _analyze_volume_drops(samples, sr)
+
+    # Composite distress score
+    result.distress_score, result.signals_detected = _compute_distress_score(result)
+
+    # Clean up temp file
+    if wav_path != audio_path and os.path.exists(wav_path):
+        os.unlink(wav_path)
+
+    # Classify distress level
+    if result.distress_score >= DISTRESS_MEDIUM:
+        distress_level = "high"
+    elif result.distress_score >= DISTRESS_LOW:
+        distress_level = "medium"
+    elif result.distress_score > 0:
+        distress_level = "low"
+    else:
+        distress_level = "none"
+
+    output = result.to_dict()
+    output["distress_level"] = distress_level
+    return output
+
+
+def get_audio_duration(audio_path: str) -> float:
+    """Get audio duration in seconds."""
+    try:
+        import librosa
+        duration = librosa.get_duration(path=audio_path)
+        return float(duration)
+    except (ImportError, Exception):
+        try:
+            import soundfile as sf
+            info = sf.info(audio_path)
+            return float(info.duration)
+        except (ImportError, Exception):
+            return 0.0