Datasets:
File size: 33,449 Bytes
cc7d399 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 | #!/usr/bin/env python3
"""
Multilingual Script-Fidelity Evaluation.
Evaluates ASR models on FLEURS test splits for ten languages and computes:
- WER, CER (standard metrics)
- Script Fidelity (SF): fraction of hypothesis characters in the target script
- Dominant script label per utterance
Output: analysis/sf_results.csv (one row per model × language)
analysis/sf_utterances/ (per-utterance JSONs for qualitative analysis)
All ten FLEURS test splits, including Pashto (ps_af), are loaded via
Hugging Face datasets.
Usage
-----
uv run python scripts/eval_multilang.py \\
--hf-token hf_xxx \\
--results-dir /workspace/results_multilang \\
--hub-repo ANONYMIZED_OUTPUT_REPO \\
--languages pashto hindi bengali malayalam somali georgian urdu arabic persian tamil \\
--whisper-sizes tiny base small medium large-v2 large-v3 turbo \\
--run-mms --run-seamless
Environment variables (fallbacks):
HF_TOKEN, RESULTS_DIR, HUB_REPO
"""
import argparse, os, sys, json, re, time, unicodedata, logging
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(Path(__file__).parent))
from runtime_cache import configure_runtime_cache
configure_runtime_cache(ROOT)
import numpy as np
import pandas as pd
import torch
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import seaborn as sns
from tqdm import tqdm
from datasets import load_dataset, Audio
from evaluate import load as load_metric
from huggingface_hub import login, HfApi
# Script fidelity module (co-located in this scripts/ directory)
from script_fidelity import (
compute_sfr, compute_sfr_batch, dominant_script, SCRIPT_CONFIGS
)
# ── LOGGING ───────────────────────────────────────────────────────────────────
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s %(levelname)s %(message)s',
handlers=[logging.StreamHandler(sys.stdout)],
)
log = logging.getLogger(__name__)
# ── ARGUMENT PARSING ──────────────────────────────────────────────────────────
def parse_args():
p = argparse.ArgumentParser(description='Multilingual script-fidelity ASR benchmark')
p.add_argument('--hf-token', default=os.environ.get('HF_TOKEN', ''))
p.add_argument('--results-dir', default=os.environ.get('RESULTS_DIR', './results_multilang'))
p.add_argument('--hub-repo', default=os.environ.get('HUB_REPO', ''))
p.add_argument('--hub-private', action='store_true')
p.add_argument('--languages', nargs='+',
default=['pashto', 'hindi', 'bengali', 'malayalam', 'somali',
'georgian', 'urdu', 'arabic', 'persian', 'tamil'],
help='Languages to evaluate')
p.add_argument('--whisper-sizes', nargs='*',
default=['tiny', 'base', 'small', 'medium', 'large-v2', 'large-v3', 'turbo'])
p.add_argument('--run-mms', action='store_true')
p.add_argument('--run-seamless', action='store_true')
p.add_argument('--run-gemma4', action='store_true',
help='Evaluate Gemma 4 E2B (requires transformers>=5.0, '
'soundfile; runs on MPS on Apple Silicon)')
p.add_argument('--sample-size', type=int, default=None,
help='Utterances per dataset per language (None = full set)')
p.add_argument('--batch-size', type=int, default=16)
p.add_argument('--force', action='store_true',
help='Rerun selected model-language pairs even if present')
return p.parse_args()
# ── HF HUB ────────────────────────────────────────────────────────────────────
def push_to_hub(results_dir: Path, hub_repo: str, api: HfApi, private: bool = False):
if not hub_repo:
return
try:
api.create_repo(hub_repo, repo_type='dataset', private=private, exist_ok=True)
api.upload_folder(
folder_path=str(results_dir),
repo_id=hub_repo,
repo_type='dataset',
commit_message='Update script-fidelity results',
ignore_patterns=['*.pyc', '__pycache__'],
)
log.info(f'Pushed results to: {hub_repo}')
except Exception as e:
log.warning(f'Hub push failed: {e}')
# ── NORMALIZATION ─────────────────────────────────────────────────────────────
# Per-language normalization: strip diacritics, punctuation, numerals.
# We normalise ONLY for WER/CER computation; the raw hypothesis is used for SF.
_DIACRITIC_RE = re.compile(r'[\u064B-\u065F\u0670]') # Arabic diacritics
_PUNCT_ARABIC = re.compile(r'[،؟؛!.«»()\[\]{}\u060C\u061B\u061F\u066D\u06D4\u200C\u200D]')
_DIGITS_ALL = re.compile(r'[0-9٠-٩]')
_TATWEEL = '\u0640'
def normalize_arabic_script(text: str) -> str:
if not isinstance(text, str): return ''
text = unicodedata.normalize('NFC', text)
text = text.replace(_TATWEEL, '')
text = _DIACRITIC_RE.sub('', text)
text = _PUNCT_ARABIC.sub('', text)
text = _DIGITS_ALL.sub('', text)
return ' '.join(text.split())
def normalize_latin(text: str) -> str:
if not isinstance(text, str): return ''
text = unicodedata.normalize('NFC', text).lower()
text = re.sub(r"[^\w\s'-]", '', text)
text = _DIGITS_ALL.sub('', text)
return ' '.join(text.split())
def normalize_indic(text: str) -> str:
if not isinstance(text, str): return ''
text = unicodedata.normalize('NFC', text)
text = re.sub(r'[।॥!?,.:;()\[\]{}"\']', '', text)
text = _DIGITS_ALL.sub('', text)
return ' '.join(text.split())
def normalize_georgian(text: str) -> str:
# Lowercase converts rare Asomtavruli capitals to standard Mkhedruli.
# \w is Unicode-aware in Python re, so Georgian letters are preserved.
if not isinstance(text, str): return ''
text = unicodedata.normalize('NFC', text).lower()
text = re.sub(r"[^\w\s'-]", '', text)
text = _DIGITS_ALL.sub('', text)
return ' '.join(text.split())
NORMALIZERS = {
'pashto': normalize_arabic_script,
'urdu': normalize_arabic_script,
'hindi': normalize_indic,
'bengali': normalize_indic,
'malayalam': normalize_indic,
'somali': normalize_latin,
'arabic': normalize_arabic_script,
'persian': normalize_arabic_script,
'tamil': normalize_indic,
'georgian': normalize_georgian,
}
# ── DATASET LOADING ───────────────────────────────────────────────────────────
def load_fleurs(language: str, sample_size: int | None) -> dict:
cfg = SCRIPT_CONFIGS[language]
code = cfg.fleurs_code
log.info(f'Loading FLEURS {code} test...')
ds = load_dataset('google/fleurs', code, split='test', trust_remote_code=True)
ds = ds.cast_column('audio', Audio(sampling_rate=16_000))
refs = [ex['transcription'] for ex in ds]
audios = [ex['audio']['array'] for ex in ds]
if sample_size:
import random; random.seed(42)
idx = sorted(random.sample(range(len(refs)), min(sample_size, len(refs))))
refs = [refs[i] for i in idx]
audios = [audios[i] for i in idx]
log.info(f' FLEURS {code}: {len(refs)} utterances')
return {'refs': refs, 'audios': audios, 'language': language, 'fleurs_code': code}
# ── RESULT STORE ──────────────────────────────────────────────────────────────
class ResultStore:
def __init__(self, csv_path: Path, force: bool = False):
self.csv_path = csv_path
self.force = force
if csv_path.exists():
self.df = pd.read_csv(csv_path)
completed_df = self.df
if 'source' in completed_df.columns:
source = completed_df['source'].fillna('')
completed_df = completed_df[source != 'paper_a_' + 'import']
self.completed = set(zip(completed_df['model'], completed_df['language']))
log.info(f'Resuming — {len(self.completed)} model×language combos done')
else:
self.df = pd.DataFrame()
self.completed = set()
def is_done(self, model_id: str, language: str) -> bool:
if self.force:
return False
return (model_id, language) in self.completed
def add(self, row: dict):
if not self.df.empty and {'model', 'language'}.issubset(self.df.columns):
same = (
(self.df['model'] == row['model'])
& (self.df['language'] == row['language'])
)
self.df = self.df.loc[~same].copy()
self.df = pd.concat([self.df, pd.DataFrame([row])], ignore_index=True)
self.df.to_csv(self.csv_path, index=False)
self.completed.add((row['model'], row['language']))
log.info(
f" Saved: {row['model']:45s} / {row['language']:12s} "
f"WER={row.get('wer_pct')}% SFR={row.get('sfr_mean')}%"
)
# ── COMPUTE METRICS ───────────────────────────────────────────────────────────
def compute_metrics(
refs: list[str],
preds: list[str],
language: str,
wer_metric,
cer_metric,
) -> dict:
norm_fn = NORMALIZERS[language]
refs_norm = [norm_fn(r) for r in refs]
pred_norm = [norm_fn(p) for p in preds]
# WER/CER — skip empty references
pairs = [(r, p) for r, p in zip(refs_norm, pred_norm) if r.strip()]
wer_val = cer_val = None
if pairs:
refs_f, preds_f = zip(*pairs)
wer_val = round(wer_metric.compute(
references=list(refs_f), predictions=list(preds_f)) * 100, 2)
cer_val = round(cer_metric.compute(
references=list(refs_f), predictions=list(preds_f)) * 100, 2)
# Script Fidelity — computed on RAW hypothesis (before normalisation)
sfr_scores, dom_scripts = compute_sfr_batch(preds, language)
sfr_valid = [s for s in sfr_scores if s is not None]
sfr_empty = sum(1 for s in sfr_scores if s is None)
sfr_mean = round(np.mean(sfr_valid) * 100, 2) if sfr_valid else None
sfr_full = sum(1 for s in sfr_valid if s == 1.0) # utterances with SFR=100%
sfr_zero = sum(1 for s in sfr_valid if s == 0.0) # script collapse utterances
dom_counts = pd.Series(dom_scripts).value_counts().to_dict()
return {
'n': len(refs),
'n_pairs': len(pairs),
'wer_pct': wer_val,
'cer_pct': cer_val,
'sfr_mean': sfr_mean,
'sfr_full_pct': round(sfr_full / max(len(sfr_valid), 1) * 100, 1),
'sfr_zero_pct': round(sfr_zero / max(len(sfr_valid), 1) * 100, 1),
'sfr_empty_n': sfr_empty,
**{f'dom_{k}': v for k, v in dom_counts.items()},
}, sfr_scores, dom_scripts
# ── WHISPER EVALUATION ────────────────────────────────────────────────────────
_WHISPER_MODEL_IDS = {
'turbo': 'openai/whisper-large-v3-turbo',
}
# Whisper language token names (from Whisper tokenizer vocabulary)
# These are the forced-language token strings, not BCP-47 codes.
WHISPER_LANG_TOKENS = {
'pashto': 'pashto',
'urdu': 'urdu',
'hindi': 'hindi',
'bengali': 'bengali',
'malayalam': 'malayalam',
'somali': 'somali',
'arabic': 'arabic',
'persian': 'persian',
'tamil': 'tamil',
'georgian': 'georgian',
}
def eval_whisper_language(
size: str,
language: str,
dataset: dict,
store: ResultStore,
preds_dir: Path,
batch_size: int,
wer_metric,
cer_metric,
device: str,
) -> None:
from transformers import pipeline as hf_pipeline
model_id = _WHISPER_MODEL_IDS.get(size, f'openai/whisper-{size}')
lang_token = WHISPER_LANG_TOKENS[language]
if store.is_done(model_id, language):
log.info(f' Skipping {model_id}/{language} (done)'); return
log.info(f' Loading {model_id} for {language}...')
pipe = hf_pipeline(
'automatic-speech-recognition',
model=model_id,
device=0 if device == 'cuda' else -1,
torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
)
refs, audios = dataset['refs'], dataset['audios']
log.info(f' {model_id} on FLEURS {language} ({len(audios)} utterances)...')
t0 = time.time()
preds = []
for i in tqdm(range(0, len(audios), batch_size), desc=f'W-{size}/{language}', leave=False):
batch = audios[i:i+batch_size]
outs = pipe(batch, batch_size=batch_size, chunk_length_s=30,
generate_kwargs={'language': lang_token, 'task': 'transcribe'})
preds.extend([o['text'] for o in outs])
elapsed = time.time() - t0
# Persist predictions
pred_file = preds_dir / f'whisper_{size}_{language}_predictions.json'
with open(pred_file, 'w', encoding='utf-8') as f:
json.dump({'model': model_id, 'language': language,
'references': refs, 'predictions': preds},
f, ensure_ascii=False)
metrics, sfr_scores, dom = compute_metrics(refs, preds, language, wer_metric, cer_metric)
total_s = sum(len(a) / 16_000 for a in audios)
store.add({
'model': model_id,
'family': 'Whisper',
'size': size,
'language': language,
'rtf': round(elapsed / total_s, 4),
**metrics,
})
del pipe
if device == 'cuda': torch.cuda.empty_cache()
# ── MMS EVALUATION ────────────────────────────────────────────────────────────
def eval_mms_language(
language: str,
dataset: dict,
store: ResultStore,
preds_dir: Path,
batch_size: int,
wer_metric,
cer_metric,
device: str,
) -> None:
from transformers import Wav2Vec2ForCTC, AutoProcessor
model_id = 'facebook/mms-1b-all'
if store.is_done(model_id, language):
log.info(f' Skipping MMS/{language} (done)'); return
cfg = SCRIPT_CONFIGS[language]
lang_codes = cfg.mms_lang
log.info(f' Loading {model_id} for {language}...')
proc = AutoProcessor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id, torch_dtype=torch.float16)
model = model.to(device).eval()
mms_lang = None
for code in lang_codes:
try:
proc.tokenizer.set_target_lang(code)
model.load_adapter(code)
mms_lang = code
log.info(f' MMS adapter: {code}')
break
except Exception as e:
log.warning(f' MMS lang={code} failed: {e}')
if mms_lang is None:
log.error(f' No MMS adapter for {language} — skipping')
del model, proc
if device == 'cuda': torch.cuda.empty_cache()
return
refs, audios = dataset['refs'], dataset['audios']
log.info(f' MMS on FLEURS {language} ({len(audios)} utterances)...')
t0 = time.time()
preds = []
for i in tqdm(range(0, len(audios), batch_size), desc=f'MMS/{language}', leave=False):
batch = audios[i:i+batch_size]
inputs = proc(batch, sampling_rate=16_000, return_tensors='pt', padding=True)
inputs = {k: (v.to(device, dtype=torch.float16) if v.is_floating_point() else v.to(device))
for k, v in inputs.items()}
with torch.no_grad():
logits = model(**inputs).logits
pred_ids = torch.argmax(logits, dim=-1)
preds.extend(proc.batch_decode(pred_ids))
elapsed = time.time() - t0
pred_file = preds_dir / f'mms_1b_{language}_predictions.json'
with open(pred_file, 'w', encoding='utf-8') as f:
json.dump({'model': model_id, 'language': language,
'references': refs, 'predictions': preds},
f, ensure_ascii=False)
metrics, _, _ = compute_metrics(refs, preds, language, wer_metric, cer_metric)
total_s = sum(len(a) / 16_000 for a in audios)
store.add({
'model': model_id,
'family': 'MMS',
'size': '1B',
'language': language,
'rtf': round(elapsed / total_s, 4),
**metrics,
})
del model, proc
if device == 'cuda': torch.cuda.empty_cache()
# ── SEAMLESSM4T EVALUATION ────────────────────────────────────────────────────
# SeamlessM4T language codes for target languages (tgt_lang parameter)
SEAMLESS_LANG_CODES = {
'pashto': 'pbt', # Southern Pashto (FLORES-200)
'urdu': 'urd',
'hindi': 'hin',
'bengali': 'ben',
'malayalam': 'mal',
'somali': 'som',
'arabic': 'arb', # Modern Standard Arabic
'persian': 'pes', # Western Persian
'tamil': 'tam',
'georgian': 'kat',
}
def eval_seamless_language(
language: str,
dataset: dict,
store: ResultStore,
preds_dir: Path,
batch_size: int,
wer_metric,
cer_metric,
device: str,
) -> None:
from transformers import AutoProcessor, SeamlessM4Tv2ForSpeechToText
model_id = 'facebook/seamless-m4t-v2-large'
tgt_lang = SEAMLESS_LANG_CODES.get(language)
if tgt_lang is None:
log.warning(f' SeamlessM4T: no lang code for {language} — skipping')
return
if store.is_done(model_id, language):
log.info(f' Skipping SeamlessM4T/{language} (done)'); return
log.info(f' Loading {model_id} (~10 GB VRAM) for {language}...')
try:
proc = AutoProcessor.from_pretrained(model_id, use_fast=False)
except Exception as e:
log.warning(f' Processor load failed ({e}); retrying with force_download...')
proc = AutoProcessor.from_pretrained(model_id, force_download=True)
try:
model = SeamlessM4Tv2ForSpeechToText.from_pretrained(
model_id, torch_dtype=torch.float16)
except Exception as e:
log.warning(f' Model load failed ({e}); retrying with force_download=True...')
model = SeamlessM4Tv2ForSpeechToText.from_pretrained(
model_id, torch_dtype=torch.float16, force_download=True)
model = model.to(device).eval()
refs, audios = dataset['refs'], dataset['audios']
s4t_batch = max(1, batch_size // 4)
log.info(f' SeamlessM4T on FLEURS {language} ({len(audios)} utterances)...')
t0 = time.time()
preds = []
for i in tqdm(range(0, len(audios), s4t_batch), desc=f'S4T/{language}', leave=False):
batch = audios[i:i+s4t_batch]
inputs = proc(audios=batch, sampling_rate=16_000,
return_tensors='pt', padding=True)
inputs = {k: (v.to(device, dtype=torch.float16) if v.is_floating_point() else v.to(device))
for k, v in inputs.items()}
with torch.no_grad():
out = model.generate(**inputs, tgt_lang=tgt_lang)
sequences = out.sequences if hasattr(out, 'sequences') else out
preds.extend(proc.batch_decode(sequences, skip_special_tokens=True))
elapsed = time.time() - t0
pred_file = preds_dir / f'seamless_m4t_v2_{language}_predictions.json'
with open(pred_file, 'w', encoding='utf-8') as f:
json.dump({'model': model_id, 'language': language,
'references': refs, 'predictions': preds},
f, ensure_ascii=False)
metrics, _, _ = compute_metrics(refs, preds, language, wer_metric, cer_metric)
total_s = sum(len(a) / 16_000 for a in audios)
store.add({
'model': model_id,
'family': 'SeamlessM4T',
'size': 'v2-large',
'language': language,
'rtf': round(elapsed / total_s, 4),
**metrics,
})
del model, proc
if device == 'cuda': torch.cuda.empty_cache()
# ── GEMMA 4 EVALUATION ───────────────────────────────────────────────────────
# Exact transcription prompt from the Gemma 4 model card.
_GEMMA4_PROMPT = (
'Transcribe the following speech segment in its original language. '
'Follow these specific instructions for formatting the answer:\n'
'* Only output the transcription, with no newlines.\n'
'* When transcribing numbers, write the digits, '
'i.e. write 1.7 and not one point seven, and write 3 instead of three.'
)
# unsloth/gemma-4-E2B-it: Apache-2.0, no HF license gate, same weights as
# google/gemma-4-E2B-it but accessible without agreeing to Google's terms.
# E2B = Effective 2B; ~10 GB BF16, ~4 GB at 4-bit. Audio supported on E2B/E4B only.
_GEMMA4_MODEL_ID = 'unsloth/gemma-4-E2B-it'
def eval_gemma4_language(
language: str,
dataset: dict,
store: ResultStore,
preds_dir: Path,
wer_metric,
cer_metric,
device: str,
) -> None:
"""Evaluate Gemma 4 E2B on one language.
Audio arrays are written to per-utterance temp WAV files because
processor.apply_chat_template only accepts file paths, not numpy arrays.
Runs at batch_size=1; Gemma 4 is an autoregressive multimodal LLM.
device_map='auto' handles placement; the device param is used for cleanup only.
"""
import tempfile
import soundfile as sf
try:
from transformers import AutoModelForMultimodalLM, AutoProcessor
except ImportError:
log.error('transformers>=5.0 required for AutoModelForMultimodalLM; skipping')
return
if store.is_done(_GEMMA4_MODEL_ID, language):
log.info(f' Skipping Gemma4/{language} (done)'); return
# E2B at dtype='auto': ~10 GB BF16 on CUDA, ~10 GB float16 on MPS.
# AutoModelForMultimodalLM is required — AutoModelForCausalLM does not
# correctly initialise the audio encoder in Gemma 4's architecture.
log.info(f' Loading {_GEMMA4_MODEL_ID} (~10 GB)...')
proc = AutoProcessor.from_pretrained(_GEMMA4_MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
_GEMMA4_MODEL_ID,
dtype='auto',
device_map='auto',
)
model.eval()
refs, audios = dataset['refs'], dataset['audios']
log.info(f' Gemma 4 on FLEURS {language} ({len(audios)} utterances)...')
t0 = time.time()
preds = []
with tempfile.TemporaryDirectory() as tmpdir:
for i, audio_array in enumerate(
tqdm(audios, desc=f'Gemma4/{language}', leave=False)):
wav_path = str(Path(tmpdir) / f'utt_{i}.wav')
sf.write(wav_path, audio_array, 16_000)
messages = [{
'role': 'user',
'content': [
{'type': 'audio', 'audio': wav_path},
{'type': 'text', 'text': _GEMMA4_PROMPT},
],
}]
try:
inputs = proc.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors='pt',
add_generation_prompt=True,
).to(model.device)
input_len = inputs['input_ids'].shape[-1]
with torch.no_grad():
out = model.generate(
**inputs,
max_new_tokens=256,
do_sample=False,
)
text = proc.decode(out[0][input_len:], skip_special_tokens=True)
preds.append(text.strip())
except Exception as exc:
log.warning(f' Gemma4 utterance {i} failed: {exc}')
preds.append('')
elapsed = time.time() - t0
pred_file = preds_dir / f'gemma4_{language}_predictions.json'
with open(pred_file, 'w', encoding='utf-8') as f:
json.dump({'model': _GEMMA4_MODEL_ID, 'language': language,
'references': refs, 'predictions': preds},
f, ensure_ascii=False)
metrics, _, _ = compute_metrics(refs, preds, language, wer_metric, cer_metric)
total_s = sum(len(a) / 16_000 for a in audios)
store.add({
'model': _GEMMA4_MODEL_ID,
'family': 'Gemma4',
'size': 'E2B',
'language': language,
'rtf': round(elapsed / total_s, 4),
**metrics,
})
del model, proc
if device == 'cuda':
torch.cuda.empty_cache()
elif device == 'mps':
torch.mps.empty_cache()
# ── FIGURES ───────────────────────────────────────────────────────────────────
_PAPER_FAMILIES = {'Whisper', 'MMS', 'SeamlessM4T', 'Gemma4'}
def _paper_rows(df: pd.DataFrame) -> pd.DataFrame:
"""Rows included in the main paper benchmark."""
if 'family' not in df.columns:
return df
return df[df['family'].isin(_PAPER_FAMILIES)].copy()
def make_sf_heatmap(store: ResultStore, figures_dir: Path) -> None:
"""Model × language SF% heatmap — the paper's main figure."""
df = _paper_rows(store.df)
if df.empty or 'sfr_mean' not in df.columns:
return
pivot = df.pivot_table(
index='model', columns='language', values='sfr_mean', aggfunc='first')
if pivot.empty:
return
# Short model names for display
short_names = (
pivot.index
.str.replace('openai/whisper-large-v3-turbo', 'Whisper turbo')
.str.replace('openai/whisper-', 'Whisper ')
.str.replace('facebook/mms-1b-all', 'MMS 1B')
.str.replace('facebook/seamless-m4t-v2-large', 'SeamlessM4T v2')
.str.replace('unsloth/gemma-4-E2B-it', 'Gemma 4 E2B')
.str.replace('google/gemma-4-E2B-it', 'Gemma 4 E2B')
)
fig, ax = plt.subplots(figsize=(max(8, len(pivot.columns) * 1.5),
max(5, len(pivot) * 0.5)))
sns.heatmap(
pivot.values,
xticklabels=pivot.columns.tolist(),
yticklabels=short_names.tolist(),
annot=True, fmt='.1f',
cmap='RdYlGn', vmin=0, vmax=100,
linewidths=0.5, ax=ax,
cbar_kws={'label': 'Script Fidelity (%)'},
)
ax.set_title('Script Fidelity (%) by Model and Language — FLEURS Test Sets')
ax.set_xlabel('Language')
ax.set_ylabel('Model')
plt.tight_layout()
out = figures_dir / 'sfr_heatmap.pdf'
fig.savefig(out, bbox_inches='tight')
plt.close(fig)
log.info(f'Saved: {out}')
def make_wer_vs_sf_scatter(store: ResultStore, figures_dir: Path) -> None:
"""WER vs SF scatter — shows decoupling cases."""
df = _paper_rows(store.df)
if df.empty or 'sfr_mean' not in df.columns:
return
languages = sorted(df['language'].unique())
n = len(languages)
ncols = 5
nrows = int(np.ceil(n / ncols))
fig, axes = plt.subplots(nrows, ncols, figsize=(17, 7.5), sharex=True, sharey=False)
axes = np.array(axes).reshape(-1)
zone_colors = {
'collapse': '#d73027',
'mixed': '#fdae61',
'high': '#1a9850',
}
for ax, lang in zip(axes, languages):
sub = df[df['language'] == lang].dropna(subset=['wer_pct', 'sfr_mean'])
if sub.empty:
ax.set_title(lang); continue
point_colors = [
zone_colors['collapse'] if v < 10 else
zone_colors['mixed'] if v <= 90 else
zone_colors['high']
for v in sub['sfr_mean']
]
ax.scatter(sub['sfr_mean'], sub['wer_pct'], color=point_colors,
s=55, zorder=5, edgecolor='black', linewidth=0.25)
ax.axvline(10, color=zone_colors['collapse'], linestyle='--', linewidth=1)
ax.axvline(90, color=zone_colors['high'], linestyle='--', linewidth=1)
ax.set_xlabel('Script Fidelity (%)')
ax.set_ylabel('WER (%)')
ax.set_title(lang.capitalize())
ax.set_xlim(-5, 105)
for ax in axes[n:]:
ax.axis('off')
handles = [
plt.Line2D([0], [0], marker='o', color='w', label='SFR < 10%',
markerfacecolor=zone_colors['collapse'], markeredgecolor='black', markersize=7),
plt.Line2D([0], [0], marker='o', color='w', label='10-90%',
markerfacecolor=zone_colors['mixed'], markeredgecolor='black', markersize=7),
plt.Line2D([0], [0], marker='o', color='w', label='> 90%',
markerfacecolor=zone_colors['high'], markeredgecolor='black', markersize=7),
]
fig.legend(handles=handles, loc='lower center', ncol=3, frameon=False)
plt.suptitle('WER vs Script Fidelity - FLEURS test sets', y=0.98)
plt.tight_layout(rect=(0, 0.05, 1, 0.95))
out = figures_dir / 'wer_vs_sfr_scatter.pdf'
fig.savefig(out, bbox_inches='tight')
plt.close(fig)
log.info(f'Saved: {out}')
# ── MAIN ──────────────────────────────────────────────────────────────────────
def main():
args = parse_args()
if torch.cuda.is_available():
device = 'cuda'
elif torch.backends.mps.is_available():
device = 'mps'
else:
device = 'cpu'
log.info(f'Device: {device}')
if device == 'cuda':
log.info(f'GPU: {torch.cuda.get_device_name(0)} '
f'VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.1f} GB')
elif device == 'mps':
log.info(f'MPS RAM available: {torch.mps.recommended_max_memory()/1e9:.1f} GB')
if args.hf_token:
login(token=args.hf_token)
api = HfApi(token=args.hf_token or None)
results_dir = Path(args.results_dir)
preds_dir = results_dir / 'predictions'
results_dir.mkdir(parents=True, exist_ok=True)
preds_dir.mkdir(exist_ok=True)
fh = logging.FileHandler(results_dir / 'eval_multilang.log')
fh.setFormatter(logging.Formatter('%(asctime)s %(levelname)s %(message)s'))
log.addHandler(fh)
wer_metric = load_metric('wer')
cer_metric = load_metric('cer')
store = ResultStore(results_dir / 'sf_results.csv', force=args.force)
# Step 1: load FLEURS for each requested language.
datasets = {}
for lang in args.languages:
if lang not in SCRIPT_CONFIGS:
log.warning(f'Unknown language: {lang} — skipping')
continue
datasets[lang] = load_fleurs(lang, args.sample_size)
# Step 2: Whisper
for size in args.whisper_sizes:
for lang, dataset in datasets.items():
eval_whisper_language(
size, lang, dataset, store, preds_dir,
args.batch_size, wer_metric, cer_metric, device)
push_to_hub(results_dir, args.hub_repo, api, args.hub_private)
# Step 3: MMS
if args.run_mms:
for lang, dataset in datasets.items():
eval_mms_language(
lang, dataset, store, preds_dir,
args.batch_size, wer_metric, cer_metric, device)
push_to_hub(results_dir, args.hub_repo, api, args.hub_private)
# Step 4: SeamlessM4T
if args.run_seamless:
for lang, dataset in datasets.items():
eval_seamless_language(
lang, dataset, store, preds_dir,
args.batch_size, wer_metric, cer_metric, device)
push_to_hub(results_dir, args.hub_repo, api, args.hub_private)
# Step 5: Gemma 4
if args.run_gemma4:
gemma_datasets = dict(datasets) # copy; already loaded above
for lang, dataset in gemma_datasets.items():
eval_gemma4_language(
lang, dataset, store, preds_dir, wer_metric, cer_metric, device)
push_to_hub(results_dir, args.hub_repo, api, args.hub_private)
# Step 6: figures
figures_dir = Path(__file__).parent.parent / 'figures'
figures_dir.mkdir(parents=True, exist_ok=True)
make_sf_heatmap(store, figures_dir)
make_wer_vs_sf_scatter(store, figures_dir)
# Final push + summary
push_to_hub(results_dir, args.hub_repo, api, args.hub_private)
log.info('=== DONE ===')
if not store.df.empty and 'sfr_mean' in store.df.columns:
pivot = store.df.pivot_table(
index='model', columns='language', values='sfr_mean', aggfunc='first')
log.info(f'\nSF% summary:\n{pivot.to_string()}')
pivot_wer = store.df.pivot_table(
index='model', columns='language', values='wer_pct', aggfunc='first')
log.info(f'\nWER% summary:\n{pivot_wer.to_string()}')
if __name__ == '__main__':
main()
|