scripts/run_final_experiment.py

"""
FINAL COMPREHENSIVE EXPERIMENT — Everything Claude knows, all at once.

Learnings applied:
  - Seed 100: best contour for most segments
  - Seed 42: best for seg_004 (EXCELLENT DTW), good for seg_002
  - temp calibration: 1.0-1.15 for std<50, 1.8-1.95 for std>60, 2.0 for std>70
  - top_p=0.7, top_k=15: tightest contour control
  - top_p=0.75, top_k=20: balanced (default)
  - tokens = duration * 12.5: perfect duration match
  - IPA lengthening (e): works on specific content words in longer sentences
  - Cross-reference: longer clips (seg_003/005) give better voice clone
  - ? mark: creates rising intonation, helped seg_005
  - ... pauses: help rhythm on long sentences
  - Continuation mode: FAILS for JP->EN
  - Full IPA: HURTS — model handles full phoneme input differently
  - CAPS: mild effect, not reliable
  - rep_penalty: no significant effect
  - Wider top_p/top_k: more std but worse contour (tradeoff)

Strategy: For each segment, run a grid of the most promising combinations.
"""
import requests, json, time, sys

BASE = "https://yellow-yaks-matter.loca.lt"

def gen(seg_id, attempt, text, temp=1.7, top_p=0.8, top_k=25, rep=1.0,
        tokens=None, seed=42, ref_seg=None):
    payload = {
        "seg_id": seg_id, "attempt": attempt, "text": text,
        "audio_temperature": temp, "audio_top_p": top_p, "audio_top_k": top_k,
        "audio_repetition_penalty": rep, "seed": seed, "mode": "generation",
    }
    if tokens: payload["tokens"] = tokens
    if ref_seg: payload["ref_seg_id"] = ref_seg
    try:
        r = requests.post(f"{BASE}/generate", json=payload, timeout=180)
        result = r.json()
        if "error" in result:
            return None
        s = result["score"]
        f = result["f0_dtw"]
        print(f"  {seg_id}-{attempt:03d} | {s:>5.1f} | DTW={f:.3f}({result['f0_dtw_rating']:>9s}) | "
              f"F0={result['generated']['f0_mean']:.0f} std={result['generated']['f0_std']:.0f} | "
              f"t={temp} p={top_p} k={top_k} s={seed}")
        return result
    except Exception as e:
        print(f"  {seg_id}-{attempt:03d} | FAIL: {str(e)[:40]}")
        return None

requests.get(f"{BASE}/health", timeout=10)
print("="*80)
print("FINAL COMPREHENSIVE EXPERIMENT")
print("="*80)

att_counter = {1: 100, 2: 100, 3: 100, 4: 100, 5: 100}  # start from attempt 100

def next_att(sid):
    att_counter[sid] += 1
    return att_counter[sid]

# ═══════════════════════════════════════════════════════════════════
# seg_001: F0=264, std=46 (calm). Best=22.9
# Known best: temp=1.1, seed=100, ref=seg_003, IPA /everyone/
# Grid: temp[1.0, 1.05, 1.1, 1.15, 1.2] x seed[42, 77, 100] x ref[1,3]
# Plus IPA variants
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*40)
print("seg_001 — COMPREHENSIVE GRID (target: F0=264, std=46)")
print("="*40)

texts_1 = [
    ("plain", "Hello everyone. Thank you all, for, joining us today."),
    ("ipa_everyone", "Hello /\u025bv\u0279iw\u028c\u02d0n/. Thank you all, for, joining us today."),
    ("ipa_both", "Hello /\u025bv\u0279iw\u028c\u02d0n/. Thank you all, for, joining us /t\u0259de\u026a\u02d0/."),
]

for tname, text in texts_1:
    for temp in [1.0, 1.08, 1.15]:
        for seed in [42, 100]:
            for ref in [1, 3]:
                a = next_att(1)
                gen(1, a, text, temp=temp, top_p=0.7, top_k=15, tokens=41, seed=seed, ref_seg=ref)
                time.sleep(1.5)

# ═══════════════════════════════════════════════════════════════════
# seg_002: F0=235, std=70 (most expressive). Best=25.9
# Hardest segment. Problem: can't get high std + good contour.
# Grid: temp[1.6, 1.7, 1.8, 1.9] x seed[42, 77, 100, 123] x text variants
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*40)
print("seg_002 — COMPREHENSIVE GRID (target: F0=235, std=70)")
print("="*40)

texts_2 = [
    ("plain", "I'm SUEYOSHI, from the Innovation Center."),
    ("dramatic", "I'm... SUEYOSHI! From the Innovation Center."),
    ("ipa_name", "I'm /su\u02d0e\u026ajo\u028a\u0283i\u02d0/, from the Innovation Center."),
]

for tname, text in texts_2:
    for temp in [1.7, 1.8, 1.9]:
        for seed in [42, 77, 100]:
            a = next_att(2)
            gen(2, a, text, temp=temp, top_p=0.75, top_k=20, tokens=25, seed=seed)
            time.sleep(1.5)

# Also try with different refs
for seed in [42, 100]:
    for ref in [4, 5]:
        a = next_att(2)
        gen(2, a, "I'm SUEYOSHI, from the Innovation Center.",
            temp=1.8, top_p=0.75, top_k=20, tokens=25, seed=seed, ref_seg=ref)
        time.sleep(1.5)

# ═══════════════════════════════════════════════════════════════════
# seg_003: F0=240, std=67. Best=20.8
# Already good with multi-IPA. Fine-tune around sweet spot.
# Grid: temp[1.85, 1.88, 1.9, 1.92, 1.95] x seed[42, 100] x IPA variants
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*40)
print("seg_003 — FINE-TUNE GRID (target: F0=240, std=67)")
print("="*40)

texts_3 = [
    ("ipa_2", "For this /s\u025b\u02d0\u0283\u0259n/... please feel free to relax, and listen to the talk, while enjoying your /l\u028c\u02d0nt\u0283/."),
    ("ipa_4", "For this /s\u025b\u02d0\u0283\u0259n/... please feel /f\u0279i\u02d0/ to /\u0279\u026a\u02d0l\u00e6ks/, and listen to the talk, while enjoying your /l\u028c\u02d0nt\u0283/."),
]

for tname, text in texts_3:
    for temp in [1.87, 1.9, 1.93]:
        for seed in [42, 100]:
            a = next_att(3)
            gen(3, a, text, temp=temp, top_p=0.75, top_k=20, tokens=62, seed=seed)
            time.sleep(1.5)

# ═══════════════════════════════════════════════════════════════════
# seg_004: F0=283, std=75. Best=19.8 (EXCELLENT!)
# Already excellent. Grid around winner: temp[1.85-1.95] x seed[42] x top_p/k
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*40)
print("seg_004 — SQUEEZE MORE (target: F0=283, std=75)")
print("="*40)

for temp in [1.85, 1.88, 1.9, 1.92, 1.95]:
    for seed in [42, 100]:
        a = next_att(4)
        gen(4, a, "Before we begin,", temp=temp, top_p=0.75, top_k=20, tokens=21, seed=seed)
        time.sleep(1.5)

# Tighter sampling with best temp
for top_p, top_k in [(0.7, 15), (0.72, 18)]:
    a = next_att(4)
    gen(4, a, "Before we begin,", temp=1.9, top_p=top_p, top_k=top_k, tokens=21, seed=42)
    time.sleep(1.5)

# ═══════════════════════════════════════════════════════════════════
# seg_005: F0=243, std=60. Best=19.9 (EXCELLENT!)
# Grid: temp[1.82-1.88] x seed[42, 100] x text[? and plain]
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*40)
print("seg_005 — SQUEEZE MORE (target: F0=243, std=60)")
print("="*40)

for temp in [1.82, 1.84, 1.85, 1.86, 1.88]:
    for seed in [42, 100]:
        a = next_att(5)
        gen(5, a, "I'd like to ask, how familiar everyone here is with the Innovation Center?",
            temp=temp, top_p=0.75, top_k=20, tokens=63, seed=seed)
        time.sleep(1.5)

# IPA + ? combo
for seed in [42, 100]:
    a = next_att(5)
    gen(5, a, "I'd like to ask, how familiar everyone here is with the /\u026an\u0259ve\u026a\u02d0\u0283\u0259n s\u025bnt\u025a/?",
        temp=1.85, top_p=0.75, top_k=20, tokens=63, seed=seed)
    time.sleep(1.5)

# ═══════════════════════════════════════════════════════════════════
# FINAL RESULTS
# ═══════════════════════════════════════════════════════════════════
print("\n" + "="*80)
print("FINAL RESULTS — ALL EXPERIMENTS")
print("="*80)

try:
    results = requests.get(f"{BASE}/results", timeout=30).json()
    for item in results:
        print(f"  seg_{item['seg_id']:03d}: {item['total_attempts']:>3d} attempts | "
              f"best={item['best_score']:.1f} (#{item['best_attempt']}, "
              f"F0 DTW={item['best_f0_dtw']:.3f})")
    print(f"\n  Total attempts: {sum(r['total_attempts'] for r in results)}")
except Exception as e:
    print(f"  {e}")