scripts/analyze_silence_sensitivity.py

#!/usr/bin/env python3
"""Measure endpoint-model sensitivity to appended trailing silence.

The report is deliberately aggregate-only.  It never serializes record IDs,
source rows, audio, transcripts, or per-example probabilities.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import math
import os
import statistics
import sys
from collections.abc import Mapping, Sequence
from pathlib import Path
from typing import Any

REPOSITORY_ROOT = Path(__file__).resolve().parents[1]
SOURCE_ROOT = REPOSITORY_ROOT / "src"
if str(SOURCE_ROOT) not in sys.path:
    sys.path.insert(0, str(SOURCE_ROOT))

TRAILING_SILENCE_MS = (0, 200, 400, 800)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--checkpoint", required=True)
    parser.add_argument("--source", required=True, help="JSON/JSONL split manifest")
    parser.add_argument(
        "--source-root",
        required=True,
        help="base directory for manifest source_file entries",
    )
    parser.add_argument("--split", required=True)
    parser.add_argument("--output", required=True)
    parser.add_argument("--batch-size", type=int, default=16)
    parser.add_argument(
        "--max-examples",
        type=int,
        help="deterministic manifest-order prefix after split filtering",
    )
    parser.add_argument(
        "--threshold",
        type=float,
        help="default: checkpoint's validation-calibrated threshold, otherwise 0.5",
    )
    parser.add_argument("--device", default="auto")
    return parser.parse_args()


def _resolve(value: str) -> Path:
    path = Path(value).expanduser()
    return path.resolve() if path.is_absolute() else (REPOSITORY_ROOT / path).resolve()


def _portable_path(path: Path) -> str:
    try:
        return path.resolve().relative_to(REPOSITORY_ROOT).as_posix()
    except ValueError:
        return path.name


def _sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def _field(record: Any, *names: str, default: Any = None) -> Any:
    for name in names:
        if isinstance(record, Mapping) and name in record:
            return record[name]
        if hasattr(record, name):
            return getattr(record, name)
    return default


def _identity_collate(records: list[Any]) -> list[Any]:
    """Keep raw records intact so every silence condition shares one decode."""

    return records


def _device(torch: Any, requested: str) -> Any:
    if requested != "auto":
        return torch.device(requested)
    if torch.cuda.is_available():
        return torch.device("cuda")
    mps = getattr(torch.backends, "mps", None)
    if mps is not None and mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")


def _percentile(values: Sequence[float], quantile: float) -> float:
    if not values:
        raise ValueError("percentiles require at least one value")
    if not 0.0 <= quantile <= 1.0:
        raise ValueError("quantile must be in [0, 1]")
    ordered = sorted(float(value) for value in values)
    position = (len(ordered) - 1) * quantile
    lower = math.floor(position)
    upper = math.ceil(position)
    if lower == upper:
        return ordered[lower]
    fraction = position - lower
    return ordered[lower] * (1.0 - fraction) + ordered[upper] * fraction


def _distribution(values: Sequence[float]) -> dict[str, float | int]:
    if not values:
        raise ValueError("distribution requires at least one value")
    finite = [float(value) for value in values]
    if any(not math.isfinite(value) for value in finite):
        raise ValueError("distribution values must be finite")
    return {
        "count": len(finite),
        "mean": statistics.fmean(finite),
        "standard_deviation": statistics.pstdev(finite),
        "minimum": min(finite),
        "p05": _percentile(finite, 0.05),
        "p50": _percentile(finite, 0.50),
        "p95": _percentile(finite, 0.95),
        "maximum": max(finite),
    }


def _probability_shift(
    labels: Sequence[int], baseline: Sequence[float], condition: Sequence[float]
) -> dict[str, Any]:
    if not labels or len(labels) != len(baseline) or len(labels) != len(condition):
        raise ValueError("paired shift inputs must have the same non-zero length")
    deltas = [
        float(candidate) - float(reference)
        for reference, candidate in zip(baseline, condition, strict=True)
    ]
    absolute = [abs(value) for value in deltas]

    def label_distribution(label: int) -> dict[str, float | int] | None:
        selected = [delta for target, delta in zip(labels, deltas, strict=True) if target == label]
        return _distribution(selected) if selected else None

    return {
        "definition": "p_end(condition) - p_end(0ms) on the same examples",
        **_distribution(deltas),
        "mean_absolute": statistics.fmean(absolute),
        "root_mean_square": math.sqrt(statistics.fmean(value * value for value in deltas)),
        "increased_count": sum(value > 0.0 for value in deltas),
        "decreased_count": sum(value < 0.0 for value in deltas),
        "unchanged_count": sum(value == 0.0 for value in deltas),
        "by_label": {
            "HOLD": label_distribution(0),
            "END": label_distribution(1),
        },
    }


def _decision_flips(
    labels: Sequence[int],
    baseline: Sequence[float],
    condition: Sequence[float],
    threshold: float,
) -> dict[str, int | float]:
    if not labels or len(labels) != len(baseline) or len(labels) != len(condition):
        raise ValueError("paired flip inputs must have the same non-zero length")
    baseline_predictions = [int(float(value) >= threshold) for value in baseline]
    condition_predictions = [int(float(value) >= threshold) for value in condition]
    pairs = list(zip(labels, baseline_predictions, condition_predictions, strict=True))
    hold_to_end = sum(before == 0 and after == 1 for _, before, after in pairs)
    end_to_hold = sum(before == 1 and after == 0 for _, before, after in pairs)
    flips = hold_to_end + end_to_hold
    return {
        "count": flips,
        "rate": flips / len(labels),
        "unchanged_count": len(labels) - flips,
        "HOLD_to_END": hold_to_end,
        "END_to_HOLD": end_to_hold,
        "false_interruptions_introduced": sum(
            target == 0 and before == 0 and after == 1 for target, before, after in pairs
        ),
        "false_interruptions_resolved": sum(
            target == 0 and before == 1 and after == 0 for target, before, after in pairs
        ),
        "missed_ends_introduced": sum(
            target == 1 and before == 1 and after == 0 for target, before, after in pairs
        ),
        "missed_ends_resolved": sum(
            target == 1 and before == 0 and after == 1 for target, before, after in pairs
        ),
    }


def _append_silence_and_pad(
    waveforms: Sequence[Any],
    silence_ms: int,
    sample_rate: int,
    max_seconds: float,
    pad_side: str,
    torch: Any,
) -> tuple[Any, Any]:
    """Append valid zeros, suffix-crop to the model window, then batch-pad."""

    if silence_ms < 0:
        raise ValueError("silence_ms cannot be negative")
    if sample_rate <= 0 or max_seconds <= 0.0:
        raise ValueError("sample_rate and max_seconds must be positive")
    if pad_side not in {"left", "right"}:
        raise ValueError("pad_side must be 'left' or 'right'")
    if not waveforms:
        raise ValueError("waveforms cannot be empty")
    target_samples = int(round(max_seconds * sample_rate))
    if target_samples < 1:
        raise ValueError("model window rounds to zero samples")
    silence_samples = int(round(silence_ms * sample_rate / 1000.0))
    prepared: list[Any] = []
    for waveform in waveforms:
        signal = torch.as_tensor(waveform, dtype=torch.float32).flatten()
        if signal.numel() == 0:
            raise ValueError("decoded audio cannot be empty")
        if silence_samples:
            signal = torch.cat((signal, signal.new_zeros(silence_samples)))
        prepared.append(signal[-target_samples:])
    lengths = torch.tensor([signal.numel() for signal in prepared], dtype=torch.long)
    padded = torch.zeros((len(prepared), target_samples), dtype=torch.float32)
    for index, signal in enumerate(prepared):
        if pad_side == "left":
            padded[index, -signal.numel() :] = signal
        else:
            padded[index, : signal.numel()] = signal
    return padded, lengths


def _validate_args(args: argparse.Namespace) -> tuple[Path, Path, Path, Path]:
    if args.batch_size < 1:
        raise SystemExit("--batch-size must be positive")
    if args.max_examples is not None and args.max_examples < 1:
        raise SystemExit("--max-examples must be positive")
    if args.threshold is not None and (
        not math.isfinite(args.threshold) or not 0.0 <= args.threshold <= 1.0
    ):
        raise SystemExit("--threshold must be finite and in [0, 1]")
    checkpoint = _resolve(args.checkpoint)
    source = _resolve(args.source)
    source_root = _resolve(args.source_root)
    output = _resolve(args.output)
    if not checkpoint.is_file():
        raise SystemExit(f"checkpoint does not exist: {checkpoint}")
    if not source.is_file() or source.suffix.lower() not in {".json", ".jsonl"}:
        raise SystemExit("--source must be an existing JSON/JSONL manifest")
    if not source_root.is_dir():
        raise SystemExit(f"--source-root is not a directory: {source_root}")
    return checkpoint, source, source_root, output


def main() -> int:
    args = parse_args()
    checkpoint_path, source_path, source_root, output_path = _validate_args(args)
    try:
        import torch
        from torch.utils.data import DataLoader
    except ImportError as exc:
        raise SystemExit("silence sensitivity analysis requires PyTorch") from exc

    from turn_detection.models import LogMelConfig, LogMelFrontend, load_model_checkpoint
    from turn_detection.training.datasets import build_record_dataloader, decode_audio
    from turn_detection.training.metrics import binary_classification_metrics

    device = _device(torch, args.device)
    try:
        model, checkpoint = load_model_checkpoint(checkpoint_path, map_location=device)
    except (OSError, RuntimeError, ValueError) as exc:
        raise SystemExit(f"cannot load checkpoint: {exc}") from exc
    model.to(device).eval()
    metadata = checkpoint.get("metadata", {})
    if not isinstance(metadata, Mapping):
        metadata = {}
    feature_config = LogMelConfig.from_mapping(metadata.get("feature_config", {}))
    frontend = LogMelFrontend(feature_config).cpu().eval()
    max_seconds = float(metadata.get("max_seconds", 8.0))
    if not math.isfinite(max_seconds) or max_seconds <= 0.0:
        raise SystemExit("checkpoint metadata has an invalid max_seconds")
    raw_threshold = (
        args.threshold if args.threshold is not None else checkpoint.get("threshold", 0.5)
    )
    try:
        threshold = float(raw_threshold)
    except (TypeError, ValueError) as exc:
        raise SystemExit("checkpoint threshold is not numeric") from exc
    if not math.isfinite(threshold) or not 0.0 <= threshold <= 1.0:
        raise SystemExit("effective threshold must be finite and in [0, 1]")

    # Reuse the canonical source routing (including lazy Parquet resolution and
    # deterministic max_examples), but keep raw records so audio is decoded once
    # and the exact same examples are evaluated under every paired condition.
    routed_loader = build_record_dataloader(
        source_path,
        split=args.split,
        frontend=frontend,
        batch_size=args.batch_size,
        max_seconds=max_seconds,
        shuffle=False,
        num_workers=0,
        token=os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN"),
        max_examples=args.max_examples,
        source_root=source_root,
    )
    raw_loader = DataLoader(
        routed_loader.dataset,
        batch_size=args.batch_size,
        shuffle=False,
        num_workers=0,
        collate_fn=_identity_collate,
    )
    labels: list[int] = []
    probabilities: dict[int, list[float]] = {silence_ms: [] for silence_ms in TRAILING_SILENCE_MS}
    with torch.inference_mode():
        for raw_records in raw_loader:
            batch_labels: list[int] = []
            waveforms: list[Any] = []
            for record in raw_records:
                raw_label = _field(record, "endpoint", "endpoint_bool", "label", default=None)
                try:
                    numeric_label = float(raw_label)
                except (TypeError, ValueError) as exc:
                    raise SystemExit(
                        "manifest contains a missing or non-numeric endpoint label"
                    ) from exc
                if not math.isfinite(numeric_label) or numeric_label not in (0.0, 1.0):
                    raise SystemExit("manifest endpoint labels must be binary (0 or 1)")
                batch_labels.append(int(numeric_label))
                try:
                    waveform = decode_audio(
                        record,
                        target_sample_rate=feature_config.sample_rate,
                        max_seconds=max_seconds,
                    )
                except (OSError, RuntimeError, TypeError, ValueError) as exc:
                    raise SystemExit(
                        "manifest must resolve to waveform-backed audio for silence analysis: "
                        f"{exc}"
                    ) from exc
                waveforms.append(waveform)
            labels.extend(batch_labels)
            for silence_ms in TRAILING_SILENCE_MS:
                padded, lengths = _append_silence_and_pad(
                    waveforms,
                    silence_ms=silence_ms,
                    sample_rate=feature_config.sample_rate,
                    max_seconds=max_seconds,
                    pad_side=feature_config.pad_side,
                    torch=torch,
                )
                features, attention_mask = frontend(padded, lengths)
                output = model(features.to(device), attention_mask.to(device))
                scores = torch.sigmoid(output.endpoint_logits).detach().cpu().tolist()
                probabilities[silence_ms].extend(float(score) for score in scores)

    if not labels:
        raise SystemExit("manifest split produced no examples")
    if args.max_examples is not None and len(labels) > args.max_examples:
        raise SystemExit("internal error: deterministic max-examples bound was exceeded")
    baseline = probabilities[0]
    conditions: list[dict[str, Any]] = []
    for silence_ms in TRAILING_SILENCE_MS:
        scores = probabilities[silence_ms]
        if len(scores) != len(labels):
            raise SystemExit("internal error: silence-condition predictions are misaligned")
        condition: dict[str, Any] = {
            "trailing_silence_ms": silence_ms,
            "probability_summary": _distribution(scores),
            "classification_metrics": binary_classification_metrics(labels, scores, threshold),
        }
        if silence_ms == 0:
            condition["relative_to_0ms"] = None
        else:
            condition["relative_to_0ms"] = {
                "probability_shift": _probability_shift(labels, baseline, scores),
                "threshold_decision_flips": _decision_flips(labels, baseline, scores, threshold),
            }
        conditions.append(condition)

    run_metadata = metadata.get("run_metadata", {})
    if not isinstance(run_metadata, Mapping):
        run_metadata = {}
    report = {
        "schema_version": 1,
        "analysis": "trailing_silence_sensitivity",
        "checkpoint": _portable_path(checkpoint_path),
        "checkpoint_sha256": _sha256(checkpoint_path),
        "model_parameter_count": sum(parameter.numel() for parameter in model.parameters()),
        "training_status": run_metadata.get("status"),
        "data_scope": metadata.get("data_scope"),
        "data_revision": metadata.get("data_revision"),
        "source_manifest": _portable_path(source_path),
        "source_manifest_sha256": _sha256(source_path),
        "split": args.split,
        "threshold": threshold,
        "sample_rate": feature_config.sample_rate,
        "model_window_seconds": max_seconds,
        "example_count": len(labels),
        "positive_count": sum(labels),
        "negative_count": len(labels) - sum(labels),
        "selection": {
            "shuffle": False,
            "num_workers": 0,
            "policy": "manifest-order prefix after split filtering",
            "max_examples": args.max_examples,
        },
        "perturbation": {
            "conditions_ms": list(TRAILING_SILENCE_MS),
            "sample_value": 0.0,
            "policy": (
                "Decode/resample the original suffix once, append valid zero-valued samples, "
                "then suffix-crop to the checkpoint window before the canonical frontend."
            ),
        },
        "conditions": conditions,
        "privacy": {
            "aggregate_only": True,
            "contains_per_example_rows": False,
            "contains_raw_identifiers": False,
            "contains_audio_or_transcripts": False,
            "note": (
                "The report contains aggregate paired statistics and file-level SHA-256 "
                "provenance only."
            ),
        },
    }
    output_path.parent.mkdir(parents=True, exist_ok=True)
    temporary = output_path.with_name(output_path.name + ".tmp")
    temporary.write_text(
        json.dumps(report, indent=2, sort_keys=True, allow_nan=False) + "\n",
        encoding="utf-8",
    )
    temporary.replace(output_path)
    print(
        json.dumps(
            {
                "example_count": len(labels),
                "output": str(output_path),
                "conditions_ms": list(TRAILING_SILENCE_MS),
            },
            indent=2,
        )
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())