"""Do the paper's audio-side findings survive a change of voice? python analysis/06_voice_study/05_replication.py python analysis/06_voice_study/05_replication.py --skip-models # fast The voice study holds the words fixed and swaps the voice. The paper's claim is not "audio ratings are high" but "audio judgments are behaviourally distinct from text": raters commit to a winner at smaller gaps, lean on length less, favour the second-presented item, and agree with the text winner at near chance. If those signatures only exist for Kokoro af_heart they are properties of one TTS voice, not of the audio modality. This script recomputes each of them per arm. Five quantities, matching Section 4 of the paper: threshold Mean per-prompt rating gap on non-tie trials (paper Fig. 1): audio 26.0 vs text 39.8 over 100 prompts. The paper's threshold text and tables pool trials instead (audio 27.9, text 41.7); 08_cross_study/01_threshold_summary.py reports that version. length Coefficient on z-scored response character length (paper Table 1; audio b = 3.52, text b = 4.92). recency Coefficient on presented-first (paper b = -2.61: first-presented rated LOWER, i.e. the second item is favoured). icc ICC(2,1) and ICC(2,k) from crossed random effects (paper audio .333 / .821, text .295 / .788). agreement Winner agreement with the ORIGINAL TEXT condition, per prompt. Every arm-level number is also recomputed for the original af_heart audio and the original text on the SAME 30 prompts. The published values cover 100 prompts, so comparing a 30-prompt arm against them would confound the voice with the stimulus subset. The matched columns are the honest comparison; the published ones are shown only for orientation. """ from __future__ import annotations import argparse import sys import warnings from pathlib import Path import numpy as np import pandas as pd warnings.filterwarnings("ignore") sys.path.insert(0, str(Path(__file__).resolve().parent)) # for lme4.py ROOT = Path(__file__).resolve().parent.parent.parent VOICE_DATA = ROOT / "data/voice_study" CONTEXT_DATA = ROOT / "data/audio_context_study" DERIVED = ROOT / "results/06_voice_study" SD = ROOT / "data/main_study" # Published values, 100 prompts. Orientation only -- see module docstring. # ICCs are variance ratios from squared random-effect SDs, matching # 02_main_study/02_reliability_icc_alpha.py calc_icc and the variances this # module's own icc() gets from lmer_varcomp. (Earlier values -- audio # 0.333/0.821, text 0.295/0.788 -- were SD ratios and are not comparable.) PAPER = { "audio": dict(threshold=27.9, length=3.52, recency=-2.61, icc1=0.312, icck=0.806), "text": dict(threshold=41.7, length=4.92, recency=-2.61, icc1=0.241, icck=0.738), } # --------------------------------------------------------------- loading ---- def load_lengths() -> pd.Series: """AudioClip -> response character length.""" s = pd.read_csv(SD / "survey_long_format.csv") return s.set_index("AudioClip")["TextLength"] def load_arms(prompts: set[str] | None) -> pd.DataFrame: """New voice arms, plus original audio and text as reference arms.""" frames = [] abl = VOICE_DATA / "long_voice_normalized.csv" if abl.exists(): frames.append(pd.read_csv(abl)) else: print(f"!! {abl.relative_to(ROOT)} missing — run build_long_tables.py") for name, f in (("ORIGINAL af_heart", "long_audio_table_normalized.csv"), ("ORIGINAL text", "long_text_table_normalized.csv")): d = pd.read_csv(SD / f) d = d[~d["IsAttentionCheck"].astype(bool)].copy() d["Arm"] = name frames.append(d) df = pd.concat(frames, ignore_index=True) df = df.dropna(subset=["Rating"]) if prompts is not None: df = df[df["PromptID"].isin(prompts)] return df # --------------------------------------------------------------- metrics ---- def decision_threshold(g: pd.DataFrame) -> tuple[float, int]: """Mean per-prompt rating gap on trials where a winner was declared. One trial contributes two rows (the two clips), so the gap is taken within (participant, prompt) and only where overallPreference is not a Tie. """ d = g[g["overallPreference"].isin(["A", "B"])] gaps = (d.groupby(["ParticipantID", "PromptID"])["Rating"] .agg(lambda x: x.max() - x.min() if len(x) == 2 else np.nan) .dropna()) if gaps.empty: return np.nan, 0 per_prompt = gaps.groupby(level="PromptID").mean() return per_prompt.mean(), len(per_prompt) def winner(g: pd.DataFrame, col: str = "Rating") -> pd.Series: """PromptID -> winning clip, or 'tie' when the two clip means are equal.""" m = g.groupby(["PromptID", "AudioClip"])[col].mean().reset_index() out = {} for pid, sub in m.groupby("PromptID"): if len(sub) < 2: continue sub = sub.sort_values(col, ascending=False) top, second = sub.iloc[0], sub.iloc[1] out[pid] = "tie" if top[col] == second[col] else top["AudioClip"] return pd.Series(out, name="winner") def fit_bias_model(g: pd.DataFrame, lengths: pd.Series): """Paper Table 1, one arm at a time. Rating ~ presented_first + char_len_z + trial_z + (1|participant) + (1|clip) Crossed random effects, so lme4. Not via pymer4: 0.9.2 cannot convert a DataFrame under rpy2 3.6, so lme4.py drives R directly. That substitution reproduces the paper's Table 1 exactly -- see validate_against_paper.py. Returns (recency_b, recency_p, length_b, length_p) or NaNs. """ from _lme4 import lmer_fixed d = g.copy() d["char_len"] = d["AudioClip"].map(lengths) d = d.dropna(subset=["char_len", "Rating", "TrialNumber"]) if d["ParticipantID"].nunique() < 3 or len(d) < 40: return (np.nan,) * 4 for src, dst in (("char_len", "char_len_z"), ("TrialNumber", "trial_z")): sd = d[src].std() d[dst] = 0.0 if not sd else (d[src] - d[src].mean()) / sd d["presented_first"] = d["PresentedFirst"].astype(int) try: fit = lmer_fixed(d, "Rating ~ presented_first + char_len_z + trial_z " "+ (1|ParticipantID) + (1|AudioClip)") pick = lambda t, c: (float(fit.loc[t, c]) if t in fit.index else np.nan) return (pick("presented_first", "estimate"), pick("presented_first", "p"), pick("char_len_z", "estimate"), pick("char_len_z", "p")) except Exception as e: # noqa: BLE001 print(f" model failed: {type(e).__name__}: {e}") return (np.nan,) * 4 def icc(g: pd.DataFrame): """ICC(2,1) and ICC(2,k) from a crossed random-effects decomposition. Uses the paper's model on raw ratings, with order and trial position as fixed effects: Rating ~ presented_first + trial_z + (1|AudioClip) + (1|ParticipantID). Variance attributable to the stimulus over stimulus + rater + residual. k is the mean number of raters per clip, so ICC(2,k) answers "how reliable is the mean of the raters we actually have". """ from _lme4 import lmer_varcomp d = g.dropna(subset=["Rating", "TrialNumber", "AudioClip"]).copy() if d["ParticipantID"].nunique() < 3 or d["AudioClip"].nunique() < 3: return np.nan, np.nan, np.nan d["pf"] = d["PresentedFirst"].astype(int) sd = d["TrialNumber"].std() d["trial_z"] = (d["TrialNumber"] - d["TrialNumber"].mean()) / (sd if sd else 1) try: rv = lmer_varcomp( d, "Rating ~ pf + trial_z + (1|AudioClip) + (1|ParticipantID)") v_stim = float(rv.get("AudioClip", np.nan)) v_rater = float(rv.get("ParticipantID", np.nan)) v_res = float(rv.get("Residual", np.nan)) k = d.groupby("AudioClip").size().mean() icc1 = v_stim / (v_stim + v_rater + v_res) icck = v_stim / (v_stim + (v_rater + v_res) / k) return icc1, icck, k except Exception as e: # noqa: BLE001 print(f" icc failed: {type(e).__name__}: {e}") return np.nan, np.nan, np.nan # ------------------------------------------------------------------ main ---- def main() -> None: ap = argparse.ArgumentParser() ap.add_argument("--skip-models", action="store_true", help="threshold and agreement only; no lme4 fits") ap.add_argument("--all-prompts", action="store_true", help="do not restrict the reference arms to the 30") args = ap.parse_args() lengths = load_lengths() sel = pd.read_csv(ROOT / "stimuli/voice_study/prompts.csv")["PromptID"] prompts = None if args.all_prompts else set(sel) if prompts: print(f"restricted to the {len(prompts)} Ablation-A prompts " f"(reference arms recomputed on the same set)") df = load_arms(prompts) text_win = winner(df[df["Arm"] == "ORIGINAL text"]) rows = [] for arm, g in df.groupby("Arm", sort=True): n_part = g["ParticipantID"].nunique() n_prompt = g["PromptID"].nunique() per_clip = g.groupby("AudioClip").size() thr, n_thr = decision_threshold(g) w = winner(g) both = w.index.intersection(text_win.index) agree = (np.nan if len(both) == 0 else float((w[both] == text_win[both]).mean())) rec_b = rec_p = len_b = len_p = i1 = ik = kbar = np.nan if not args.skip_models: print(f" fitting {arm} ...") rec_b, rec_p, len_b, len_p = fit_bias_model(g, lengths) i1, ik, kbar = icc(g) rows.append(dict(arm=arm, n=len(g), participants=n_part, prompts=n_prompt, raters_per_clip=f"{per_clip.min()}-{per_clip.max()}", threshold=thr, recency_b=rec_b, recency_p=rec_p, length_b=len_b, length_p=len_p, icc1=i1, icck=ik, k=kbar, agree_with_text=agree, n_agree=len(both))) res = pd.DataFrame(rows) out = DERIVED / "voice_replication.csv" res.to_csv(out, index=False) pd.set_option("display.width", 200, "display.max_columns", 50) print("\n" + "=" * 78) print("REPLICATION OF THE PAPER'S AUDIO SIGNATURES, BY VOICE") print("=" * 78) show = res[["arm", "participants", "prompts", "raters_per_clip", "threshold", "recency_b", "length_b", "icc1", "icck", "agree_with_text"]] print(show.to_string(index=False, float_format=lambda x: f"{x:7.3f}")) print("\npublished (100 prompts, for orientation only):") for k, v in PAPER.items(): print(f" {k:6} threshold {v['threshold']:.1f} length b {v['length']:.2f}" f" recency b {v['recency']:.2f} ICC(2,1) {v['icc1']:.3f}" f" ICC(2,k) {v['icck']:.3f}") thin = res[res["participants"] < 10] if not thin.empty: print("\n!! UNDERPOWERED — fewer than 10 raters in: " + ", ".join(thin["arm"])) print(" The design targets 10 ratings per item. Treat every number " "above as a pipeline check, not a result.") print(f"\nwrote {out.relative_to(ROOT)}") if __name__ == "__main__": sys.exit(main())