Datasets:
File size: 14,248 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 | #!/usr/bin/env python3
"""
Script Fidelity Rate (SFR) metric for multilingual ASR evaluation.
SFR = fraction of non-whitespace, non-punctuation characters in the ASR hypothesis
that belong to the expected target script block(s).
A value of 1.0 means every output character is in the correct script.
A value near 0 means the model produced output entirely in the wrong script —
a condition termed *script collapse* (e.g. Latin transliteration or Arabic when
Devanagari was expected).
SFR is reference-free: it requires only the hypothesis string and a target
language identifier, not a ground-truth transcription. This makes it usable
as a deployment-time audit metric even when no labelled data is available.
This generalises a Pashto-specific script-check heuristic to ten target
languages and scripts.
Usage
-----
>>> from script_fidelity import compute_sfr, SCRIPT_CONFIGS
>>> sfr = compute_sfr("کابل کې ښه هوا ده", "pashto")
>>> assert sfr == 1.0
>>> sfr = compute_sfr("this is entirely Latin", "pashto")
>>> assert sfr == 0.0
"""
import re
import unicodedata
from dataclasses import dataclass, field
from typing import Optional
# ── UNICODE BLOCK RANGES ──────────────────────────────────────────────────────
# Each language config lists one or more (lo, hi) inclusive code-point ranges
# that define the "target script" for that language. An output character is
# in-script if it falls in ANY of the target ranges OR is in the optional
# unique_codepoints set.
@dataclass
class ScriptConfig:
"""Configuration for a single target script."""
name: str # human-readable label
ranges: list[tuple[int, int]] # [(lo, hi), ...] inclusive
unique_codepoints: set[str] = field(default_factory=set)
# Optional: codepoints that CONFIRM the script even if below 70% threshold.
# Useful for Pashto which shares Arabic script with Urdu/Dari but has
# language-unique glyphs.
fleurs_code: str = '' # google/fleurs dataset code
mms_lang: list[str] = field(default_factory=list) # MMS language adapter IDs
# Pashto-unique codepoints (not found in standard Arabic or Urdu):
# ټ U+0679 ډ U+0688 ڼ U+06BC ړ U+0693 ښ U+069A ږ U+0696 ځ U+0681
# ۍ U+06CD ګ U+06AB ې U+06D0 ۀ U+06C0 څ U+0685
_PASHTO_UNIQUE = set('ټډڼړښږځۍګېۀڅ')
SCRIPT_CONFIGS: dict[str, ScriptConfig] = {
# Pashto — Perso-Arabic + Pashto-unique glyphs
'pashto': ScriptConfig(
name='Pashto (Perso-Arabic)',
ranges=[
(0x0600, 0x06FF), # Arabic block (covers all Perso-Arabic letters)
(0x0750, 0x077F), # Arabic Supplement
(0xFB50, 0xFDFF), # Arabic Presentation Forms-A
(0xFE70, 0xFEFF), # Arabic Presentation Forms-B
],
unique_codepoints=_PASHTO_UNIQUE,
fleurs_code='ps_af',
mms_lang=['pbu', 'pbt', 'pus'],
),
# Urdu — Perso-Arabic (shares blocks with Pashto; no unique codepoints beyond Urdu-specific)
# Urdu-unique: ں U+06BA ٹ U+0679 ڈ U+0688 ڑ U+0691 ے U+06D2
'urdu': ScriptConfig(
name='Urdu (Perso-Arabic)',
ranges=[
(0x0600, 0x06FF),
(0x0750, 0x077F),
(0xFB50, 0xFDFF),
(0xFE70, 0xFEFF),
],
unique_codepoints=set('ںٹڈڑے'),
fleurs_code='ur_pk',
# MMS 1B: no bare 'urd'; Urdu adapters are urd-script_* (not 'udu', a different lang).
mms_lang=['urd-script_arabic', 'urd-script_devanagari', 'urd-script_latin'],
),
# Hindi — Devanagari
'hindi': ScriptConfig(
name='Hindi (Devanagari)',
ranges=[
(0x0900, 0x097F), # Devanagari
(0xA8E0, 0xA8FF), # Devanagari Extended
],
fleurs_code='hi_in',
mms_lang=['hin'],
),
# Bengali — Bengali script
'bengali': ScriptConfig(
name='Bengali (Bengali script)',
ranges=[
(0x0980, 0x09FF), # Bengali
],
fleurs_code='bn_in',
mms_lang=['ben'],
),
# Malayalam — Malayalam script
'malayalam': ScriptConfig(
name='Malayalam (Malayalam script)',
ranges=[
(0x0D00, 0x0D7F), # Malayalam
],
fleurs_code='ml_in',
mms_lang=['mal'],
),
# Somali — Latin script (historically used Arabic; modern standard = Latin)
# Somali uses basic Latin + no diacritics in standard orthography; the
# main failure mode for Whisper is outputting Arabic on Somali audio.
# Two separate ranges for A-Z and a-z: a single 0x41-0x7A range would
# include ^ (U+005E) and ` (U+0060), both category Sk, which pass
# _is_countable and would be falsely counted as in-script.
'somali': ScriptConfig(
name='Somali (Latin)',
ranges=[
(0x0041, 0x005A), # Basic Latin A–Z
(0x0061, 0x007A), # Basic Latin a–z
(0x00C0, 0x024F), # Latin Extended-A and Extended-B
],
fleurs_code='so_so',
mms_lang=['som'],
),
# Arabic (Modern Standard Arabic) — same script family as Pashto/Urdu
# Added as a control: MSA models should produce correct Arabic script.
# Interesting case: does Whisper confuse Arabic audio with Urdu/Pashto?
'arabic': ScriptConfig(
name='Arabic (Modern Standard Arabic)',
ranges=[
(0x0600, 0x06FF),
(0x0750, 0x077F),
(0xFB50, 0xFDFF),
(0xFE70, 0xFEFF),
],
fleurs_code='ar_eg',
mms_lang=['ara'], # MMS uses ISO 639-3; SeamlessM4T uses 'arb' (FLORES-200)
),
# Persian/Farsi — Perso-Arabic script with Persian-specific letters
# پ U+067E چ U+0686 ژ U+0698 گ U+06AF are unique to Persian vs Arabic
'persian': ScriptConfig(
name='Persian/Farsi (Perso-Arabic)',
ranges=[
(0x0600, 0x06FF),
(0x0750, 0x077F),
(0xFB50, 0xFDFF),
(0xFE70, 0xFEFF),
],
unique_codepoints=set('پچژگ'),
fleurs_code='fa_ir',
mms_lang=['fas'], # MMS uses ISO 639-3; SeamlessM4T uses 'pes' (FLORES-200)
),
# Tamil — Tamil script
'tamil': ScriptConfig(
name='Tamil',
ranges=[
(0x0B80, 0x0BFF), # Tamil block
],
fleurs_code='ta_in',
mms_lang=['tam'],
),
# Georgian — Mkhedruli script (unique, no Latin/Arabic overlap)
# Georgian is a strong positive control: script collapse would be easy to detect.
'georgian': ScriptConfig(
name='Georgian (Mkhedruli)',
ranges=[
(0x10A0, 0x10FF), # Georgian (Asomtavruli + Mkhedruli)
(0x2D00, 0x2D2F), # Georgian Supplement (Nuskhuri)
(0x1C90, 0x1CBF), # Georgian Extended (Mtavruli capitals)
],
fleurs_code='ka_ge',
mms_lang=['kat'],
),
}
# ── HELPER FUNCTIONS ──────────────────────────────────────────────────────────
def _is_in_range(cp: int, ranges: list[tuple[int, int]]) -> bool:
return any(lo <= cp <= hi for lo, hi in ranges)
def _is_countable(ch: str) -> bool:
"""Return True for characters that should count toward the SFR denominator.
Whitespace, punctuation (Unicode category P*), and combining marks are
excluded so that diacritics and punctuation do not artificially inflate or
deflate the metric.
"""
cat = unicodedata.category(ch)
return (
not ch.isspace()
and not cat.startswith('P') # punctuation
and not cat.startswith('Z') # separators
and not cat.startswith('C') # control / format chars
)
def compute_sfr(
text: str,
language: str,
config: Optional[ScriptConfig] = None,
) -> float:
"""Compute Script Fidelity Rate for a single ASR hypothesis string.
SFR is reference-free: no ground-truth transcription is needed.
It can be computed in production deployments to detect script collapse
without any labelled data.
Parameters
----------
text : str
Raw ASR hypothesis (unnormalized is fine; NFC is applied internally).
language : str
Key into SCRIPT_CONFIGS, e.g. 'pashto', 'hindi', 'somali'.
Ignored if `config` is supplied directly.
config : ScriptConfig, optional
Use a custom config instead of looking up `language`.
Returns
-------
float
Fraction in [0.0, 1.0]. Returns ``None`` if the text has no countable
characters (empty / whitespace-only / punctuation-only hypothesis).
A value near 0 indicates script collapse.
"""
if config is None:
if language not in SCRIPT_CONFIGS:
raise ValueError(
f"Unknown language '{language}'. "
f"Available: {sorted(SCRIPT_CONFIGS)}"
)
config = SCRIPT_CONFIGS[language]
text = unicodedata.normalize('NFC', text) if text else ''
chars = [ch for ch in text if _is_countable(ch)]
if not chars:
return None # hypothesis is empty / only punctuation
in_script = sum(
1 for ch in chars
if (ch in config.unique_codepoints)
or _is_in_range(ord(ch), config.ranges)
)
return in_script / len(chars)
def dominant_script(text: str) -> str:
"""Classify the dominant script of a text string.
Returns one of: 'pashto', 'arabic_dari_urdu', 'devanagari', 'bengali',
'malayalam', 'latin', 'empty', or 'other'.
This is a fast heuristic for tallying script distributions across a corpus.
For the SF metric, use compute_sf() with a specific language config.
"""
if not text or not text.strip():
return 'empty'
# Pashto-unique glyphs confirm Pashto unambiguously
if any(ch in _PASHTO_UNIQUE for ch in text):
return 'pashto'
chars = [ch for ch in text if _is_countable(ch)]
if not chars:
return 'empty'
counts: dict[str, int] = {
'arabic_dari_urdu': 0,
'devanagari': 0,
'bengali': 0,
'malayalam': 0,
'tamil': 0,
'georgian': 0,
'latin': 0,
'other': 0,
}
for ch in chars:
cp = ord(ch)
if 0x0600 <= cp <= 0x06FF or 0xFB50 <= cp <= 0xFDFF or 0xFE70 <= cp <= 0xFEFF:
counts['arabic_dari_urdu'] += 1
elif 0x0900 <= cp <= 0x097F or 0xA8E0 <= cp <= 0xA8FF:
counts['devanagari'] += 1
elif 0x0980 <= cp <= 0x09FF:
counts['bengali'] += 1
elif 0x0D00 <= cp <= 0x0D7F:
counts['malayalam'] += 1
elif 0x0B80 <= cp <= 0x0BFF:
counts['tamil'] += 1
elif (0x10A0 <= cp <= 0x10FF) or (0x2D00 <= cp <= 0x2D2F) or (0x1C90 <= cp <= 0x1CBF):
counts['georgian'] += 1
elif (0x0041 <= cp <= 0x007A) or (0x00C0 <= cp <= 0x024F):
counts['latin'] += 1
else:
counts['other'] += 1
total = len(chars)
best = max(counts, key=counts.__getitem__)
if counts[best] / total >= 0.5:
return best
return 'other'
def compute_sfr_batch(
texts: list[str],
language: str,
) -> tuple[list[Optional[float]], list[str]]:
"""Vectorised version of compute_sfr + dominant_script.
Returns
-------
sfr_scores : list of float | None
dom_scripts : list of str
"""
config = SCRIPT_CONFIGS[language]
sfr_scores = [compute_sfr(t, language, config) for t in texts]
dom = [dominant_script(t) for t in texts]
return sfr_scores, dom
# Backward-compatibility aliases
compute_sf = compute_sfr
compute_sf_batch = compute_sfr_batch
# ── VALIDATION ────────────────────────────────────────────────────────────────
def _validate_pashto_calibration() -> None:
"""Smoke-test the Pashto SFR implementation.
Checks that Pashto-unique detection works on a known positive and a known
negative.
"""
ps_text = 'کابل کې ښه هوا ده' # contains ښ (U+069A), Pashto-unique
lat_text = 'this is entirely latin output'
sfr_ps = compute_sfr(ps_text, 'pashto')
sfr_lat = compute_sfr(lat_text, 'pashto')
assert sfr_ps == 1.0, f'Expected SFR=1.0 for Pashto text, got {sfr_ps}'
assert sfr_lat == 0.0, f'Expected SFR=0.0 for Latin text against Pashto config, got {sfr_lat}'
dom_ps = dominant_script(ps_text)
dom_lat = dominant_script(lat_text)
assert dom_ps == 'pashto', f'dominant_script failed for Pashto: {dom_ps}'
assert dom_lat == 'latin', f'dominant_script failed for Latin: {dom_lat}'
print('Pashto calibration: PASS')
def _validate_devanagari() -> None:
hi_text = 'नमस्ते' # Hindi Devanagari
lat_text = 'namaste'
sfr_hi = compute_sfr(hi_text, 'hindi')
sfr_lat = compute_sfr(lat_text, 'hindi')
assert sfr_hi == 1.0, f'Expected SFR=1.0 for Hindi, got {sfr_hi}'
assert sfr_lat == 0.0, f'Expected SFR=0.0 for Latin vs Hindi config, got {sfr_lat}'
print('Hindi (Devanagari) calibration: PASS')
def _validate_somali() -> None:
so_text = 'Somali waa luuqad' # basic Latin
ar_text = 'كابل في هواء جيد' # Arabic
sfr_so = compute_sfr(so_text, 'somali')
sfr_ar = compute_sfr(ar_text, 'somali')
assert sfr_so == 1.0, f'Expected SFR=1.0 for Somali Latin, got {sfr_so}'
assert sfr_ar == 0.0, f'Expected SFR=0.0 for Arabic vs Somali config, got {sfr_ar}'
print('Somali (Latin) calibration: PASS')
if __name__ == '__main__':
_validate_pashto_calibration()
_validate_devanagari()
_validate_somali()
# Print config summary
print('\nScript configurations:')
for lang, cfg in SCRIPT_CONFIGS.items():
n_ranges = len(cfg.ranges)
n_unique = len(cfg.unique_codepoints)
print(f' {lang:12s} {cfg.name:35s} ranges={n_ranges} unique_codepoints={n_unique}')
|