Spaces:
Running on Zero
Running on Zero
File size: 51,903 Bytes
076e1f2 36d9a65 076e1f2 424ad4c 076e1f2 e935730 7d43283 076e1f2 424ad4c 076e1f2 a3a12d5 076e1f2 cf482bc 076e1f2 398f643 076e1f2 a3a12d5 424ad4c 203078f 424ad4c 400c600 398f643 a3a12d5 398f643 076e1f2 424ad4c 9ea8036 36d9a65 076e1f2 a3a12d5 076e1f2 9ea8036 076e1f2 a3a12d5 076e1f2 a3a12d5 076e1f2 b5947dc 076e1f2 a3a12d5 424ad4c 203078f 9ea8036 398f643 400c600 398f643 9ea8036 a3a12d5 4df694e 076e1f2 36d9a65 076e1f2 36d9a65 076e1f2 a3a12d5 398f643 a3a12d5 398f643 a3a12d5 e935730 424ad4c 076e1f2 424ad4c 076e1f2 e935730 076e1f2 a3a12d5 076e1f2 a3a12d5 076e1f2 a3a12d5 076e1f2 398f643 076e1f2 398f643 076e1f2 a3a12d5 076e1f2 424ad4c 203078f 424ad4c 203078f 9ea8036 398f643 9ea8036 203078f 424ad4c 400c600 203078f 400c600 203078f 424ad4c 9ea8036 424ad4c 398f643 4df694e 3cb1117 de9ca55 a3a12d5 de9ca55 a3a12d5 de9ca55 a3a12d5 de9ca55 fb0efb7 de9ca55 7d43283 a3a12d5 076e1f2 a3a12d5 3cb1117 398f643 4df694e e935730 076e1f2 424ad4c 076e1f2 e935730 076e1f2 e935730 076e1f2 e935730 076e1f2 8aaed61 a3a12d5 de9ca55 a3a12d5 076e1f2 424ad4c 398f643 de9ca55 a3a12d5 424ad4c 398f643 076e1f2 36d9a65 7d43283 a3a12d5 7d43283 a3a12d5 7d43283 a3a12d5 424ad4c a3a12d5 7d43283 a3a12d5 7d43283 36d9a65 bc45982 cf482bc 36d9a65 bc45982 4df694e bc45982 36d9a65 a3a12d5 36d9a65 a3a12d5 076e1f2 36d9a65 076e1f2 36d9a65 076e1f2 36d9a65 916a339 bc45982 076e1f2 a3a12d5 36d9a65 a3a12d5 076e1f2 36d9a65 076e1f2 a3a12d5 424ad4c a3a12d5 076e1f2 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 36d9a65 bc45982 a3a12d5 36d9a65 a3a12d5 36d9a65 076e1f2 7d43283 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 | from __future__ import annotations
import base64
import io
import json
import math
import os
import re
import tempfile
import traceback
from typing import Any
from pathlib import Path
# Unsloth's compiled Gemma 4 audio path can trip TorchDynamo on ZeroGPU's
# runtime torch build. Keep inference eager for reliability.
os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
os.environ.setdefault("UNSLOTH_COMPILE_DISABLE", "1")
import gradio as gr
import librosa
import numpy as np
import soundfile as sf
import spaces
import torch
from transformers import AutoProcessor, Gemma4ForConditionalGeneration
from live_audio_policy import (
LiveAudioPolicy,
compute_live_audio_diagnostics,
decide_live_analysis,
validate_live_audio_diagnostics,
)
ALLOWED_CLASSES = {"clear", "frontal", "lateral", "dental", "palatal"}
DEFAULT_MODEL_ID = "thomasjvu/lisper-gemma4-e2b-audio-full"
DEFAULT_ADAPTER_ID = ""
SPACE_ROOT = Path(__file__).resolve().parent
ACOUSTIC_MODEL_PATH = SPACE_ROOT / "acoustic_model.json"
ACOUSTIC_EXTRATREES_MODEL_PATH = SPACE_ROOT / "acoustic_extratrees_v18.joblib"
ACOUSTIC_K = 5
ACOUSTIC_MIN_CONFIDENCE = 0.42
KNN_OVERRIDE_MAX_DISTANCE = 0.25
KNN_OVERRIDE_MIN_CONFIDENCE = 0.90
LIVE_CLEAR_MIN_CONFIDENCE = 0.85
LIVE_CLEAR_MIN_MARGIN = 0.25
LIVE_NONCLEAR_MIN_CONFIDENCE = 0.55
LIVE_NONCLEAR_MIN_MARGIN = 0.12
MIN_AUDIO_SECONDS = 0.45
MIN_AUDIO_RMS = 0.0015
MIN_AUDIO_PEAK = 0.012
MIN_VOICED_RATIO = 0.002
MIN_SPEECH_FRAME_RATIO = 0.04
MIN_TONAL_FRAME_RATIO = 0.04
MIN_SIBILANT_FRAME_RATIO = 0.015
MAX_NOISE_FLATNESS = 0.40
MAX_CLIPPING_RATIO = 0.08
DEFAULT_PROMPT = """Analyze this pronunciation attempt for lisp type and give concise corrective coaching.
Return exactly four labeled lines in this order:
Detected class: clear|frontal|lateral|dental|palatal
Reason: one brief reason tied to tongue placement or airflow
Corrective cue: one concrete next-step cue
Encouragement: one brief supportive line"""
CLASS_TEMPLATES = {
"clear": {
"reason": "The acoustic pattern did not strongly match the trained lisp-pattern examples, so this is treated as a tentative clear result.",
"cue": "Repeat once at a relaxed pace and keep the airflow centered through the front of the mouth.",
},
"dental": {
"reason": "The acoustic pattern is closest to the dental examples, where tongue contact near the teeth can narrow the /s/ groove.",
"cue": "Relax the tongue slightly off the teeth and keep a narrow stream of air moving forward.",
},
"frontal": {
"reason": "The acoustic pattern is closest to the frontal examples, which often sound like the airflow is too far forward.",
"cue": "Keep the tongue tip just behind the upper teeth and send the air straight forward through a small groove.",
},
"lateral": {
"reason": "The acoustic pattern is closest to the lateral examples, where air may be escaping around the sides of the tongue.",
"cue": "Start from a light /t/ position, seal the tongue sides, and let the air move forward through the center.",
},
"palatal": {
"reason": "The acoustic pattern is closest to the palatal examples, where the tongue can sit too far back and muffle the sound.",
"cue": "Bring the tongue tip slightly forward behind the upper teeth and brighten the airflow.",
},
}
GUARDED_CLASS_TEMPLATES = {
"dental": {
"reason": "The acoustic model was not confident enough to call this clear; the nearest non-clear pattern is dental, which can happen when the tongue presses too close to the teeth.",
"cue": "Try one slower repetition with the tongue relaxed just behind the teeth and the air moving forward through a narrow center groove.",
},
"frontal": {
"reason": "The acoustic model was not confident enough to call this clear; the nearest non-clear pattern is frontal, which can happen when the tongue or airflow moves too far forward.",
"cue": "Keep the tongue tip behind the upper teeth and avoid letting it push between the teeth during /s/ sounds.",
},
"lateral": {
"reason": "The acoustic model was not confident enough to call this clear; the nearest non-clear pattern is lateral, where air may be leaking around the tongue sides.",
"cue": "Seal the tongue sides lightly against the upper molars and aim the air straight down the middle.",
},
"palatal": {
"reason": "The acoustic model was not confident enough to call this clear; the nearest non-clear pattern is palatal, where the tongue may be sitting too far back.",
"cue": "Bring the tongue tip forward just behind the upper teeth and brighten the /s/ airflow.",
},
}
BROWSER_RECORDER_START_JS = r"""
async (payload) => {
const state = window.__lisperRecorder || {};
if (state.recording) {
return [payload || "", "Recording is already active.", ""];
}
if (!navigator.mediaDevices || !navigator.mediaDevices.getUserMedia) {
return [payload || "", "This browser cannot access microphone recording.", ""];
}
const stream = await navigator.mediaDevices.getUserMedia({
audio: {
channelCount: 1,
echoCancellation: false,
noiseSuppression: false,
autoGainControl: true,
},
video: false,
});
const AudioContextCtor = window.AudioContext || window.webkitAudioContext;
const audioContext = new AudioContextCtor({ sampleRate: 16000 });
await audioContext.resume();
const source = audioContext.createMediaStreamSource(stream);
const processor = audioContext.createScriptProcessor(4096, 1, 1);
const silentGain = audioContext.createGain();
silentGain.gain.value = 0;
const chunks = [];
let peak = 0;
let sumSquares = 0;
let sampleCount = 0;
processor.onaudioprocess = (event) => {
if (!window.__lisperRecorder?.recording) {
return;
}
const input = event.inputBuffer.getChannelData(0);
const copy = new Float32Array(input.length);
copy.set(input);
chunks.push(copy);
for (let i = 0; i < copy.length; i += 1) {
const value = copy[i];
const absValue = Math.abs(value);
if (absValue > peak) peak = absValue;
sumSquares += value * value;
}
sampleCount += copy.length;
};
source.connect(processor);
processor.connect(silentGain);
silentGain.connect(audioContext.destination);
window.__lisperRecorder = {
recording: true,
stream,
audioContext,
source,
processor,
silentGain,
chunks,
startedAt: Date.now(),
getStats: () => ({
peak,
rms: sampleCount ? Math.sqrt(sumSquares / sampleCount) : 0,
sampleCount,
sampleRate: audioContext.sampleRate,
}),
};
return ["", "Recording through Web Audio... press Stop when finished.", ""];
}
"""
BROWSER_RECORDER_STOP_JS = r"""
async () => {
const state = window.__lisperRecorder;
if (!state || !state.recording) {
return ["", "No active browser recording. Press Record first.", ""];
}
state.recording = false;
try { state.processor.disconnect(); } catch (_) {}
try { state.source.disconnect(); } catch (_) {}
try { state.silentGain.disconnect(); } catch (_) {}
for (const track of state.stream.getTracks()) {
track.stop();
}
const stats = state.getStats();
const sampleRate = stats.sampleRate || 16000;
const totalLength = state.chunks.reduce((sum, chunk) => sum + chunk.length, 0);
const samples = new Float32Array(totalLength);
let offset = 0;
for (const chunk of state.chunks) {
samples.set(chunk, offset);
offset += chunk.length;
}
await state.audioContext.close().catch(() => undefined);
window.__lisperRecorder = null;
function writeString(view, byteOffset, string) {
for (let i = 0; i < string.length; i += 1) {
view.setUint8(byteOffset + i, string.charCodeAt(i));
}
}
function encodeWav(floatSamples, wavSampleRate) {
const bytesPerSample = 2;
const blockAlign = bytesPerSample;
const buffer = new ArrayBuffer(44 + floatSamples.length * bytesPerSample);
const view = new DataView(buffer);
writeString(view, 0, "RIFF");
view.setUint32(4, 36 + floatSamples.length * bytesPerSample, true);
writeString(view, 8, "WAVE");
writeString(view, 12, "fmt ");
view.setUint32(16, 16, true);
view.setUint16(20, 1, true);
view.setUint16(22, 1, true);
view.setUint32(24, wavSampleRate, true);
view.setUint32(28, wavSampleRate * blockAlign, true);
view.setUint16(32, blockAlign, true);
view.setUint16(34, 16, true);
writeString(view, 36, "data");
view.setUint32(40, floatSamples.length * bytesPerSample, true);
let byteOffset = 44;
for (let i = 0; i < floatSamples.length; i += 1, byteOffset += 2) {
const clamped = Math.max(-1, Math.min(1, floatSamples[i]));
view.setInt16(byteOffset, clamped < 0 ? clamped * 0x8000 : clamped * 0x7fff, true);
}
return new Blob([view], { type: "audio/wav" });
}
const blob = encodeWav(samples, sampleRate);
const dataUrl = await new Promise((resolve, reject) => {
const reader = new FileReader();
reader.onload = () => resolve(reader.result);
reader.onerror = () => reject(reader.error);
reader.readAsDataURL(blob);
});
const durationSeconds = sampleRate ? samples.length / sampleRate : 0;
const payload = JSON.stringify({
source: "browser-web-audio-wav",
data_url: dataUrl,
mime_type: "audio/wav",
sample_rate: sampleRate,
sample_count: samples.length,
duration_seconds: Number(durationSeconds.toFixed(3)),
peak: Number(stats.peak.toFixed(6)),
rms: Number(stats.rms.toFixed(6)),
created_at: new Date().toISOString(),
});
const status = stats.peak < 0.003
? `Clip captured but appears very quiet. peak=${stats.peak.toFixed(6)} rms=${stats.rms.toFixed(6)}. Check browser microphone permission/input.`
: `Clip ready: ${durationSeconds.toFixed(1)}s, peak=${stats.peak.toFixed(3)}. Playback should contain your voice.`;
const playback = `<audio controls src="${dataUrl}" style="width:100%;"></audio>`;
return [payload, status, playback];
}
"""
BROWSER_RECORDER_CLEAR_JS = r"""
async () => {
const state = window.__lisperRecorder;
if (state?.recording) {
state.recording = false;
try { state.processor.disconnect(); } catch (_) {}
try { state.source.disconnect(); } catch (_) {}
try { state.silentGain.disconnect(); } catch (_) {}
for (const track of state.stream.getTracks()) {
track.stop();
}
await state.audioContext.close().catch(() => undefined);
}
window.__lisperRecorder = null;
return ["", "No browser recording ready.", ""];
}
"""
class InvalidAudioError(ValueError):
"""Raised when a clip is too short or too quiet to analyze honestly."""
def __init__(self, message: str, diagnostics: dict[str, Any]):
super().__init__(message)
self.diagnostics = diagnostics
def env_int(name: str, default: int) -> int:
try:
return int(os.environ.get(name, str(default)))
except ValueError:
return default
def env_float(name: str, default: float) -> float:
try:
return float(os.environ.get(name, str(default)))
except ValueError:
return default
def model_id() -> str:
return os.environ.get("LISPER_ZERO_GPU_MODEL_ID", DEFAULT_MODEL_ID).strip() or DEFAULT_MODEL_ID
def adapter_id() -> str:
return os.environ.get("LISPER_ZERO_GPU_ADAPTER_ID", DEFAULT_ADAPTER_ID).strip()
def max_new_tokens() -> int:
return env_int("LISPER_ZERO_GPU_MAX_NEW_TOKENS", 96)
def max_seq_length() -> int:
return env_int("LISPER_ZERO_GPU_MAX_SEQ_LENGTH", 2048)
def zero_gpu_size() -> str:
requested = os.environ.get("LISPER_ZERO_GPU_SIZE", "large").strip().lower()
return "xlarge" if requested == "xlarge" else "large"
def eager_load_enabled() -> bool:
return os.environ.get("LISPER_ZERO_GPU_EAGER_LOAD", "0").strip() != "0"
def load_in_4bit_enabled() -> bool:
default = "1" if adapter_id() else "0"
return os.environ.get("LISPER_ZERO_GPU_LOAD_IN_4BIT", default).strip() != "0"
def acoustic_hint_enabled() -> bool:
return os.environ.get("LISPER_ZERO_GPU_ACOUSTIC_HINT", "1").strip() != "0"
def acoustic_model_preference() -> str:
requested = os.environ.get("LISPER_ZERO_GPU_ACOUSTIC_MODEL", "auto").strip().lower()
if requested in {"extratrees", "knn"}:
return requested
return "auto"
def live_clear_min_confidence() -> float:
return env_float("LISPER_ZERO_GPU_LIVE_CLEAR_MIN_CONFIDENCE", LIVE_CLEAR_MIN_CONFIDENCE)
def live_clear_min_margin() -> float:
return env_float("LISPER_ZERO_GPU_LIVE_CLEAR_MIN_MARGIN", LIVE_CLEAR_MIN_MARGIN)
def live_nonclear_min_confidence() -> float:
return env_float(
"LISPER_ZERO_GPU_LIVE_NONCLEAR_MIN_CONFIDENCE",
env_float("LISPER_ZERO_GPU_LIVE_NONCLEAR_MIN_SCORE", LIVE_NONCLEAR_MIN_CONFIDENCE),
)
def live_nonclear_min_margin() -> float:
return env_float("LISPER_ZERO_GPU_LIVE_NONCLEAR_MIN_MARGIN", LIVE_NONCLEAR_MIN_MARGIN)
def knn_override_max_distance() -> float:
return env_float("LISPER_ZERO_GPU_KNN_OVERRIDE_MAX_DISTANCE", KNN_OVERRIDE_MAX_DISTANCE)
def knn_override_min_confidence() -> float:
return env_float("LISPER_ZERO_GPU_KNN_OVERRIDE_MIN_CONFIDENCE", KNN_OVERRIDE_MIN_CONFIDENCE)
def live_audio_policy() -> LiveAudioPolicy:
return LiveAudioPolicy(
min_audio_seconds=env_float("LISPER_ZERO_GPU_MIN_AUDIO_SECONDS", MIN_AUDIO_SECONDS),
min_peak=env_float("LISPER_ZERO_GPU_MIN_AUDIO_PEAK", MIN_AUDIO_PEAK),
min_rms=env_float("LISPER_ZERO_GPU_MIN_AUDIO_RMS", MIN_AUDIO_RMS),
min_voiced_ratio=env_float("LISPER_ZERO_GPU_MIN_VOICED_RATIO", MIN_VOICED_RATIO),
min_speech_frame_ratio=env_float("LISPER_ZERO_GPU_MIN_SPEECH_FRAME_RATIO", MIN_SPEECH_FRAME_RATIO),
min_tonal_frame_ratio=env_float("LISPER_ZERO_GPU_MIN_TONAL_FRAME_RATIO", MIN_TONAL_FRAME_RATIO),
min_sibilant_frame_ratio=env_float("LISPER_ZERO_GPU_MIN_SIBILANT_FRAME_RATIO", MIN_SIBILANT_FRAME_RATIO),
max_noise_flatness=env_float("LISPER_ZERO_GPU_MAX_NOISE_FLATNESS", MAX_NOISE_FLATNESS),
max_clipping_ratio=env_float("LISPER_ZERO_GPU_MAX_CLIPPING_RATIO", MAX_CLIPPING_RATIO),
clear_min_confidence=live_clear_min_confidence(),
clear_min_margin=live_clear_min_margin(),
nonclear_min_confidence=live_nonclear_min_confidence(),
nonclear_min_margin=live_nonclear_min_margin(),
)
def audio_alignment_enabled() -> bool:
default = "0" if adapter_id() else "1"
return os.environ.get("LISPER_ZERO_GPU_ALIGN_AUDIO_TOKENS", default).strip() != "0"
def gemma_generation_enabled() -> bool:
return os.environ.get("LISPER_ZERO_GPU_USE_GEMMA_GENERATION", "0").strip() != "0"
def torch_dtype() -> torch.dtype:
requested = os.environ.get("LISPER_ZERO_GPU_DTYPE", "float16").strip().lower()
if requested == "bfloat16":
return torch.bfloat16
if requested == "float32":
return torch.float32
return torch.float16
def auth_token() -> str | None:
token = os.environ.get("HF_TOKEN", "").strip()
return token or None
def decode_browser_audio_payload(payload: str) -> np.ndarray:
try:
parsed = json.loads(payload)
data_url = str(parsed.get("data_url") or "")
if "," not in data_url:
raise ValueError("Browser recorder payload is missing audio data.")
_, encoded = data_url.split(",", 1)
audio_bytes = base64.b64decode(encoded)
waveform, sample_rate = sf.read(io.BytesIO(audio_bytes), dtype="float32", always_2d=False)
except Exception as exc:
raise InvalidAudioError(
f"Browser recording could not be decoded: {type(exc).__name__}: {exc}",
{"status": "invalid_browser_audio_payload"},
) from exc
if waveform.ndim > 1:
waveform = waveform.mean(axis=1)
if sample_rate != 16000:
waveform = librosa.resample(waveform, orig_sr=sample_rate, target_sr=16000)
return waveform.astype(np.float32)
def normalize_audio(audio_value: str | tuple[int, np.ndarray] | None) -> np.ndarray:
if audio_value is None:
raise gr.Error("Record or upload a short audio clip first.")
if isinstance(audio_value, str):
if audio_value.strip().startswith("{"):
waveform = decode_browser_audio_payload(audio_value)
sample_rate = 16000
else:
waveform, sample_rate = sf.read(audio_value, dtype="float32", always_2d=False)
else:
sample_rate, waveform = audio_value
waveform = np.asarray(waveform, dtype=np.float32)
if waveform.ndim > 1:
waveform = waveform.mean(axis=1)
if sample_rate != 16000:
waveform = librosa.resample(waveform, orig_sr=sample_rate, target_sr=16000)
# Keep ZeroGPU requests bounded.
max_samples = 12 * 16000
if waveform.shape[0] > max_samples:
waveform = waveform[:max_samples]
peak = float(np.max(np.abs(waveform))) if waveform.size else 0.0
if peak > 1.0:
waveform = waveform / peak
return waveform.astype(np.float32)
def audio_diagnostics(waveform: np.ndarray, sample_rate: int = 16000) -> dict[str, Any]:
return compute_live_audio_diagnostics(waveform, sample_rate=sample_rate, policy=live_audio_policy())
def validate_audio_for_analysis(waveform: np.ndarray) -> dict[str, Any]:
diagnostics = audio_diagnostics(waveform)
decision = validate_live_audio_diagnostics(diagnostics, live_audio_policy())
if decision["status"] != "accepted":
diagnostics["live_audio_gate"] = decision
raise InvalidAudioError(str(decision["reason"]), diagnostics)
return diagnostics
def write_temp_audio(waveform: np.ndarray) -> str:
fd, path = tempfile.mkstemp(prefix="lisper-zero-gpu-", suffix=".wav")
os.close(fd)
sf.write(path, waveform, 16000)
return path
def build_messages(target_text: str, audio_url: str, acoustic_result: dict[str, Any] | None = None) -> list[dict[str, Any]]:
instruction = DEFAULT_PROMPT
if target_text.strip():
instruction += f'\n\nTarget text: "{target_text.strip()}"'
if acoustic_result:
instruction += (
"\n\nAcoustic pre-analysis from the waveform: "
f"class={acoustic_result['detected_class']}, "
f"confidence={acoustic_result['confidence']:.3f}. "
"Use this exact class for the Detected class line. Do not override it."
)
return [
{
"role": "system",
"content": [
{
"type": "text",
"text": "You are Lisper, a supportive speech-therapy assistant focused on concise lisp coaching.",
}
],
},
{
"role": "user",
"content": [
{"type": "audio", "url": audio_url},
{"type": "text", "text": instruction},
],
},
]
def build_runtime() -> tuple[Any, Any]:
repo_id = model_id()
adapter_repo_id = adapter_id()
token = auth_token()
if adapter_repo_id:
import unsloth # noqa: F401
from unsloth import FastVisionModel
kwargs = {
"model_name": adapter_repo_id,
"max_seq_length": max_seq_length(),
"load_in_4bit": load_in_4bit_enabled(),
"full_finetuning": False,
}
if token:
kwargs["token"] = token
model, processor = FastVisionModel.from_pretrained(**kwargs)
FastVisionModel.for_inference(model)
model.eval()
return processor, model
processor_source = adapter_repo_id or repo_id
processor = AutoProcessor.from_pretrained(processor_source, token=token, trust_remote_code=True)
model = Gemma4ForConditionalGeneration.from_pretrained(
repo_id,
token=token,
torch_dtype=torch_dtype(),
device_map={"": "cuda"},
trust_remote_code=True,
)
model.eval()
return processor, model
RUNTIME: tuple[Any, Any] | None = build_runtime() if eager_load_enabled() else None
LAST_INPUT_SUMMARY: dict[str, Any] = {}
def load_runtime() -> tuple[Any, Any]:
global RUNTIME
if RUNTIME is None:
RUNTIME = build_runtime()
return RUNTIME
def strip_generation_artifacts(text: str) -> str:
return text.replace("```", "").replace("<bos>", "").strip()
def extract_line(label: str, text: str) -> str:
match = re.search(rf"^{label}:\s*(.+)$", text, flags=re.IGNORECASE | re.MULTILINE)
return match.group(1).strip() if match else ""
def normalize_label(text: str) -> str:
value = text.strip().lower()
if "inconclusive" in value or "unclear" in value:
return "inconclusive"
for candidate in ALLOWED_CLASSES:
if candidate in value:
return candidate
return "inconclusive"
def parse_response(response: str) -> dict[str, Any]:
detected = normalize_label(extract_line("Detected class", response))
return {
"detected_class": detected,
"reason": extract_line("Reason", response),
"corrective_cue": extract_line("Corrective cue", response),
"encouragement": extract_line("Encouragement", response),
"raw_response": response,
"model_id": model_id(),
"adapter_id": adapter_id() or None,
}
def acoustic_normalize_audio(audio: np.ndarray) -> np.ndarray:
audio = np.asarray(audio, dtype=np.float32).reshape(-1)
if audio.size == 0:
return audio
audio = audio - float(np.mean(audio))
peak = float(np.max(np.abs(audio)))
if peak > 0:
audio = audio * (0.98 / peak)
return audio.astype(np.float32)
def frame_audio(audio: np.ndarray, sr: int, frame_ms: float = 25.0, hop_ms: float = 10.0) -> np.ndarray:
frame = max(1, int(sr * frame_ms / 1000))
hop = max(1, int(sr * hop_ms / 1000))
if len(audio) < frame:
audio = np.pad(audio, (0, frame - len(audio)))
count = 1 + (len(audio) - frame) // hop
shape = (count, frame)
strides = (audio.strides[0] * hop, audio.strides[0])
return np.lib.stride_tricks.as_strided(audio, shape=shape, strides=strides).copy()
def summarize_feature_values(values: np.ndarray) -> list[float]:
values = np.asarray(values, dtype=np.float64)
values = values[np.isfinite(values)]
if values.size == 0:
return [0.0] * 6
return [
float(np.mean(values)),
float(np.std(values)),
float(np.min(values)),
float(np.max(values)),
float(np.percentile(values, 10)),
float(np.percentile(values, 90)),
]
def extract_acoustic_features(audio: np.ndarray, sr: int = 16000) -> np.ndarray:
if audio.size == 0:
return np.zeros(88, dtype=np.float32)
audio = acoustic_normalize_audio(audio)
frames = frame_audio(audio, sr)
window = np.hanning(frames.shape[1]).astype(np.float32)
spectra = np.abs(np.fft.rfft(frames * window, axis=1)).astype(np.float64)
freqs = np.fft.rfftfreq(frames.shape[1], d=1.0 / sr).astype(np.float64)
power = spectra**2
eps = 1e-10
total = power.sum(axis=1) + eps
centroid = (power * freqs).sum(axis=1) / total
bandwidth = np.sqrt((power * (freqs[None, :] - centroid[:, None]) ** 2).sum(axis=1) / total)
cumulative = np.cumsum(power, axis=1)
rolloff_idx = np.argmax(cumulative >= 0.85 * total[:, None], axis=1)
rolloff = freqs[rolloff_idx]
flatness = np.exp(np.mean(np.log(power + eps), axis=1)) / (np.mean(power + eps, axis=1))
zcr = np.mean(np.abs(np.diff(np.signbit(frames), axis=1)), axis=1)
rms = np.sqrt(np.mean(frames**2, axis=1) + eps)
entropy = -(power / total[:, None] * np.log((power / total[:, None]) + eps)).sum(axis=1) / math.log(
power.shape[1]
)
def band_ratio(low: float, high: float) -> np.ndarray:
mask = (freqs >= low) & (freqs < high)
if not np.any(mask):
return np.zeros(power.shape[0])
return power[:, mask].sum(axis=1) / total
bands = [
band_ratio(0, 800),
band_ratio(800, 1800),
band_ratio(1800, 3200),
band_ratio(3200, 5000),
band_ratio(5000, min(7900, sr / 2)),
band_ratio(3500, min(7500, sr / 2)),
]
deltas = np.diff(centroid, prepend=centroid[0])
features: list[float] = [
float(len(audio) / sr),
float(np.mean(audio)),
float(np.std(audio)),
float(np.max(np.abs(audio))),
]
for values in [centroid, bandwidth, rolloff, flatness, zcr, rms, entropy, deltas, *bands]:
features.extend(summarize_feature_values(values))
return np.asarray(features, dtype=np.float32)
ACOUSTIC_MODEL: dict[str, Any] | None = None
ACOUSTIC_EXTRATREES_MODEL: dict[str, Any] | None = None
def load_acoustic_model() -> dict[str, Any] | None:
global ACOUSTIC_MODEL
if not acoustic_hint_enabled():
return None
if ACOUSTIC_MODEL is None:
if not ACOUSTIC_MODEL_PATH.exists():
return None
ACOUSTIC_MODEL = json.loads(ACOUSTIC_MODEL_PATH.read_text(encoding="utf-8"))
return ACOUSTIC_MODEL
def load_acoustic_extratrees_model() -> dict[str, Any] | None:
global ACOUSTIC_EXTRATREES_MODEL
if not acoustic_hint_enabled():
return None
if not ACOUSTIC_EXTRATREES_MODEL_PATH.exists():
return None
if ACOUSTIC_EXTRATREES_MODEL is None:
import joblib
ACOUSTIC_EXTRATREES_MODEL = joblib.load(ACOUSTIC_EXTRATREES_MODEL_PATH)
return ACOUSTIC_EXTRATREES_MODEL
def classify_acoustic_extratrees(waveform: np.ndarray) -> dict[str, Any] | None:
model = load_acoustic_extratrees_model()
if model is None:
return None
features = extract_acoustic_features(waveform, sr=int(model.get("sample_rate", 16000))).reshape(1, -1)
classifier = model["classifier"]
prediction = str(classifier.predict(features)[0])
confidence = 1.0
class_scores: dict[str, float] = {}
if hasattr(classifier, "predict_proba"):
probabilities = classifier.predict_proba(features)[0]
classes = [str(label) for label in classifier.classes_]
class_scores = {
label: round(float(probability), 6)
for label, probability in sorted(zip(classes, probabilities), key=lambda item: item[1], reverse=True)
}
confidence = float(class_scores.get(prediction, 0.0))
return {
"detected_class": prediction,
"raw_class": prediction,
"confidence": confidence,
"class_scores": class_scores,
"model_name": model.get("name", "lisper_v18_extratrees_acoustic_hint"),
"train_rows": model.get("train_rows"),
"feature_count": model.get("feature_count"),
"holdout_accuracy": model.get("holdout_accuracy"),
"low_confidence_defaulted_to_clear": False,
}
def apply_live_clear_guard(result: dict[str, Any] | None) -> dict[str, Any] | None:
if result is not None:
result["live_clear_guard_applied"] = False
return result
def classify_acoustic_knn(waveform: np.ndarray) -> dict[str, Any] | None:
model = load_acoustic_model()
if model is None:
return None
features = extract_acoustic_features(waveform, sr=int(model.get("sample_rate", 16000)))
mean = np.asarray(model["mean"], dtype=np.float32)
std = np.asarray(model["std"], dtype=np.float32)
normalized = (features - mean) / np.where(std < 1e-6, 1.0, std)
distances = []
for exemplar in model["exemplars"]:
exemplar_features = np.asarray(exemplar["features"], dtype=np.float32)
distance = float(np.linalg.norm(normalized - exemplar_features))
distances.append((distance, exemplar["label"], exemplar.get("source_id", "")))
distances.sort(key=lambda item: item[0])
class_scores: dict[str, float] = {label: 0.0 for label in model["classes"]}
for distance, label, _source_id in distances[:ACOUSTIC_K]:
class_scores[label] += 1.0 / max(distance, 1e-4)
ranked = sorted(class_scores.items(), key=lambda item: item[1], reverse=True)
top_label, top_score = ranked[0]
total_score = sum(class_scores.values()) or 1.0
confidence = float(top_score / total_score)
detected_class = top_label if confidence >= ACOUSTIC_MIN_CONFIDENCE else "clear"
return {
"detected_class": detected_class,
"raw_class": top_label,
"confidence": confidence,
"nearest_distance": round(distances[0][0], 4),
"nearest_source_id": distances[0][2],
"class_scores": {label: round(float(score), 6) for label, score in ranked},
"model_name": model.get("name"),
"low_confidence_defaulted_to_clear": detected_class == "clear" and top_label != "clear",
}
def _compact_acoustic_result(result: dict[str, Any] | None) -> dict[str, Any] | None:
if result is None:
return None
keys = (
"detected_class",
"raw_class",
"confidence",
"nearest_distance",
"nearest_source_id",
"class_scores",
"model_name",
"low_confidence_defaulted_to_clear",
)
return {key: result[key] for key in keys if key in result}
def maybe_apply_knn_override(
extratrees_result: dict[str, Any] | None,
knn_result: dict[str, Any] | None,
) -> dict[str, Any] | None:
if extratrees_result is None or knn_result is None:
return extratrees_result
knn_label = normalize_label(str(knn_result.get("raw_class") or knn_result.get("detected_class") or ""))
if knn_label == "clear":
return extratrees_result
confidence = float(knn_result.get("confidence") or 0.0)
nearest_distance = float(knn_result.get("nearest_distance") or math.inf)
max_distance = knn_override_max_distance()
min_confidence = knn_override_min_confidence()
if confidence < min_confidence or nearest_distance > max_distance:
return {
**extratrees_result,
"hybrid_override_applied": False,
"hybrid_override_reason": "knn_not_close_enough",
"hybrid_override_thresholds": {
"max_distance": max_distance,
"min_confidence": min_confidence,
},
"knn_result": _compact_acoustic_result(knn_result),
}
class_scores = {
label: float(score)
for label, score in (knn_result.get("class_scores") or {}).items()
if normalize_label(str(label)) in ALLOWED_CLASSES
}
return {
**extratrees_result,
"detected_class": knn_label,
"raw_class": knn_label,
"confidence": confidence,
"class_scores": class_scores,
"model_name": "lisper_hybrid_extratrees_knn_synthetic_override",
"low_confidence_defaulted_to_clear": False,
"hybrid_override_applied": True,
"hybrid_override_reason": "knn_close_synthetic_exemplar",
"hybrid_override_thresholds": {
"max_distance": max_distance,
"min_confidence": min_confidence,
},
"extratrees_result": _compact_acoustic_result(extratrees_result),
"knn_result": _compact_acoustic_result(knn_result),
}
def classify_acoustic(waveform: np.ndarray) -> dict[str, Any] | None:
preference = acoustic_model_preference()
if preference == "extratrees":
return classify_acoustic_extratrees(waveform)
if preference == "knn":
return classify_acoustic_knn(waveform)
extratrees_result = classify_acoustic_extratrees(waveform)
knn_result = classify_acoustic_knn(waveform)
if extratrees_result is not None:
return maybe_apply_knn_override(extratrees_result, knn_result)
return knn_result
def enforce_acoustic_response(response: str, acoustic_result: dict[str, Any] | None) -> tuple[str, dict[str, Any]]:
parsed = parse_response(response)
if not acoustic_result:
return response, parsed
detected_class = normalize_label(str(acoustic_result["detected_class"]))
if acoustic_result.get("live_clear_guard_applied"):
template = GUARDED_CLASS_TEMPLATES.get(detected_class, CLASS_TEMPLATES[detected_class])
else:
template = CLASS_TEMPLATES[detected_class]
encouragement = parsed.get("encouragement") or "Good effort. One focused repetition is enough for the next try."
final_response = "\n".join(
[
f"Detected class: {detected_class}",
f"Reason: {template['reason']}",
f"Corrective cue: {template['cue']}",
f"Encouragement: {encouragement}",
]
)
final_parsed = parse_response(final_response)
final_parsed["raw_model_response"] = response
final_parsed["acoustic_hint_enforced"] = True
return final_response, final_parsed
def build_inconclusive_response(
decision: dict[str, Any],
acoustic_result: dict[str, Any] | None,
clip_diagnostics: dict[str, Any],
) -> tuple[str, dict[str, Any]]:
reason = str(decision.get("decision_reason") or "The clip was not reliable enough to classify.")
if decision.get("status") == "error":
response = "Analysis unavailable. The acoustic model is not loaded, so Lisper will not guess a class."
else:
response = "\n".join(
[
"Detected class: inconclusive",
f"Reason: {reason}",
"Corrective cue: Record one clear phrase with /s/ or /z/ sounds, close to the microphone, then try again.",
"Encouragement: The clip was captured; we just need a cleaner attempt before giving a label.",
]
)
parsed = parse_response(response)
parsed["status"] = str(decision.get("status") or "inconclusive")
parsed["raw_model_response"] = None
parsed["acoustic_hint_enforced"] = False
parsed["audio_diagnostics"] = clip_diagnostics
parsed["acoustic_analysis"] = acoustic_result
parsed["live_audio_gate"] = decision
return response, parsed
def build_detected_acoustic_response(
acoustic_result: dict[str, Any],
decision: dict[str, Any],
clip_diagnostics: dict[str, Any],
) -> tuple[str, dict[str, Any]]:
detected_class = normalize_label(str(decision.get("detected_class") or acoustic_result.get("detected_class") or ""))
if detected_class not in CLASS_TEMPLATES:
detected_class = "inconclusive"
if detected_class == "inconclusive":
return build_inconclusive_response(
{
**decision,
"status": "inconclusive",
"decision_reason": "The live gate did not produce a valid class label.",
},
acoustic_result,
clip_diagnostics,
)
template = CLASS_TEMPLATES[detected_class]
response = "\n".join(
[
f"Detected class: {detected_class}",
f"Reason: {template['reason']}",
f"Corrective cue: {template['cue']}",
"Encouragement: Nice work getting a usable recording. Try one focused repetition next.",
]
)
parsed = parse_response(response)
parsed["status"] = "detected"
parsed["raw_model_response"] = None
parsed["acoustic_hint_enforced"] = True
parsed["gemma_generation_skipped"] = True
parsed["audio_diagnostics"] = clip_diagnostics
parsed["acoustic_analysis"] = acoustic_result
parsed["live_audio_gate"] = decision
return response, parsed
def build_audio_only_inconclusive_decision(clip_diagnostics: dict[str, Any]) -> dict[str, Any] | None:
policy = live_audio_policy()
if clip_diagnostics["sibilant_frame_ratio"] >= policy.min_sibilant_frame_ratio:
return None
return {
"status": "inconclusive",
"detected_class": "inconclusive",
"candidate_class": None,
"decision_reason": "The clip has speech energy, but not enough usable /s/ or /z/ airflow evidence.",
"thresholds": {
"min_audio_seconds": policy.min_audio_seconds,
"min_peak": policy.min_peak,
"min_rms": policy.min_rms,
"min_voiced_ratio": policy.min_voiced_ratio,
"min_speech_frame_ratio": policy.min_speech_frame_ratio,
"min_tonal_frame_ratio": policy.min_tonal_frame_ratio,
"min_sibilant_frame_ratio": policy.min_sibilant_frame_ratio,
"max_noise_flatness": policy.max_noise_flatness,
"max_clipping_ratio": policy.max_clipping_ratio,
"clear_min_confidence": policy.clear_min_confidence,
"clear_min_margin": policy.clear_min_margin,
"nonclear_min_confidence": policy.nonclear_min_confidence,
"nonclear_min_margin": policy.nonclear_min_margin,
},
"audio_diagnostics": clip_diagnostics,
"classifier": {"available": False, "skipped": "insufficient_sibilant_evidence"},
}
def audio_token_id(processor: Any) -> int | None:
value = getattr(processor, "audio_token_id", None)
if value is not None:
return int(value)
tokenizer = getattr(processor, "tokenizer", None)
token = getattr(processor, "audio_token", None) or getattr(tokenizer, "audio_token", None)
if tokenizer is not None and token is not None:
return int(tokenizer.convert_tokens_to_ids(token))
return None
def replace_audio_token_run(input_ids: torch.Tensor, token_id: int, count: int) -> tuple[torch.Tensor, dict[str, int]]:
positions = (input_ids == token_id).nonzero(as_tuple=False).flatten()
if positions.numel() == 0:
return input_ids, {"original_audio_tokens": 0, "aligned_audio_tokens": count}
start = int(positions[0].item())
end = start
while end < input_ids.shape[0] and int(input_ids[end].item()) == token_id:
end += 1
replacement = torch.full((count,), token_id, dtype=input_ids.dtype, device=input_ids.device)
aligned = torch.cat([input_ids[:start], replacement, input_ids[end:]], dim=0)
return aligned, {"original_audio_tokens": end - start, "aligned_audio_tokens": count}
def model_inference_dtype(model: Any) -> torch.dtype:
dtype = getattr(model, "dtype", None)
if dtype is not None:
return dtype
base_model = getattr(model, "base_model", None)
dtype = getattr(base_model, "dtype", None)
return dtype or torch_dtype()
def module_parameter_dtype(module: Any) -> torch.dtype:
try:
return next(module.parameters()).dtype
except StopIteration:
return model_inference_dtype(module)
def audio_input_dtype(model: Any) -> torch.dtype:
requested = os.environ.get("LISPER_ZERO_GPU_AUDIO_DTYPE", "").strip().lower()
if requested == "float16":
return torch.float16
if requested == "float32":
return torch.float32
if requested == "bfloat16" or adapter_id():
return torch.bfloat16
return module_parameter_dtype(audio_feature_module(model))
def audio_feature_module(model: Any) -> Any:
candidates = [
model,
getattr(model, "model", None),
getattr(model, "base_model", None),
getattr(getattr(model, "base_model", None), "model", None),
getattr(getattr(getattr(model, "base_model", None), "model", None), "model", None),
]
for candidate in candidates:
if candidate is not None and hasattr(candidate, "get_audio_features"):
return candidate
raise AttributeError("Could not locate Gemma audio feature module on loaded model.")
def summarize_inputs(inputs: Any) -> dict[str, Any]:
summary: dict[str, Any] = {}
for key, value in dict(inputs).items():
if hasattr(value, "shape") and hasattr(value, "dtype"):
summary[key] = {
"shape": [int(dim) for dim in value.shape],
"dtype": str(value.dtype),
"device": str(getattr(value, "device", "")),
}
else:
summary[key] = {"type": type(value).__name__}
return summary
def align_audio_placeholders(inputs: Any, processor: Any, model: Any) -> tuple[Any, dict[str, int]]:
if not audio_alignment_enabled():
return inputs, {"audio_alignment_skipped": 1}
if "input_features" not in inputs or "input_features_mask" not in inputs:
return inputs, {}
token_id = audio_token_id(processor)
if token_id is None:
return inputs, {}
with torch.inference_mode():
audio_output = audio_feature_module(model).get_audio_features(
inputs["input_features"],
inputs["input_features_mask"],
return_dict=True,
)
encoded_count = int(audio_output.attention_mask.sum().item())
if encoded_count <= 0 or inputs["input_ids"].shape[0] != 1:
return inputs, {"encoded_audio_tokens": encoded_count}
new_input_ids, metadata = replace_audio_token_run(inputs["input_ids"][0], token_id, encoded_count)
metadata["encoded_audio_tokens"] = encoded_count
if metadata["original_audio_tokens"] == encoded_count:
return inputs, metadata
inputs["input_ids"] = new_input_ids.unsqueeze(0)
inputs["attention_mask"] = torch.ones_like(inputs["input_ids"])
if "mm_token_type_ids" in inputs and hasattr(processor, "create_mm_token_type_ids"):
mm_token_type_ids = processor.create_mm_token_type_ids(inputs["input_ids"].detach().cpu())
inputs["mm_token_type_ids"] = torch.as_tensor(
mm_token_type_ids,
dtype=inputs["input_ids"].dtype,
device=inputs["input_ids"].device,
)
return inputs, metadata
def _analyze_impl(audio: str | tuple[int, np.ndarray] | None, target_text: str) -> tuple[str, str]:
global LAST_INPUT_SUMMARY
waveform = normalize_audio(audio)
clip_diagnostics = validate_audio_for_analysis(waveform)
audio_only_decision = build_audio_only_inconclusive_decision(clip_diagnostics)
if audio_only_decision is not None:
response, parsed = build_inconclusive_response(audio_only_decision, None, clip_diagnostics)
return response, json.dumps(parsed, indent=2)
acoustic_result = classify_acoustic(waveform)
live_decision = decide_live_analysis(acoustic_result, clip_diagnostics, live_audio_policy())
if live_decision["status"] != "detected":
response, parsed = build_inconclusive_response(live_decision, acoustic_result, clip_diagnostics)
return response, json.dumps(parsed, indent=2)
if not gemma_generation_enabled():
response, parsed = build_detected_acoustic_response(acoustic_result, live_decision, clip_diagnostics)
return response, json.dumps(parsed, indent=2)
audio_url = write_temp_audio(waveform)
processor, model = load_runtime()
messages = build_messages(target_text, audio_url, acoustic_result)
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
enable_thinking=False,
return_dict=True,
return_tensors="pt",
)
device = next(model.parameters()).device
if hasattr(inputs, "to"):
inputs = inputs.to(device)
for float_key in ("input_features", "pixel_values", "pixel_values_videos"):
if float_key in inputs and hasattr(inputs[float_key], "to"):
inputs[float_key] = inputs[float_key].to(dtype=model_inference_dtype(model))
if "input_features" in inputs:
audio_dtype = audio_input_dtype(model)
inputs["input_features"] = inputs["input_features"].to(dtype=audio_dtype)
if "input_features_mask" in inputs and hasattr(inputs["input_features_mask"], "to"):
inputs["input_features_mask"] = inputs["input_features_mask"].to(device=device)
inputs, alignment = align_audio_placeholders(inputs, processor, model)
LAST_INPUT_SUMMARY = summarize_inputs(inputs)
with torch.inference_mode():
outputs = model.generate(
**inputs,
max_new_tokens=max_new_tokens(),
do_sample=False,
use_cache=True,
)
prompt_length = inputs["input_ids"].shape[1]
decoded = processor.decode(outputs[0][prompt_length:], skip_special_tokens=True)
raw_response = strip_generation_artifacts(decoded)
response, parsed = enforce_acoustic_response(raw_response, acoustic_result)
parsed["status"] = "detected"
parsed["audio_token_alignment"] = alignment
parsed["audio_diagnostics"] = clip_diagnostics
parsed["acoustic_analysis"] = acoustic_result
parsed["live_audio_gate"] = live_decision
return response, json.dumps(parsed, indent=2)
def analyze_with_errors(audio: str | tuple[int, np.ndarray] | None, target_text: str) -> tuple[str, str]:
try:
return _analyze_impl(audio, target_text)
except InvalidAudioError as exc:
payload = {
"status": "rejected_audio",
"reason": str(exc),
"audio_diagnostics": exc.diagnostics,
}
return (
"Recording not usable yet. Please record a clear speech clip before analysis.",
json.dumps(payload, indent=2),
)
except Exception as exc:
payload = {
"error_type": type(exc).__name__,
"message": str(exc),
"model_id": model_id(),
"adapter_id": adapter_id() or None,
"dtype": os.environ.get("LISPER_ZERO_GPU_DTYPE", "float16"),
"load_in_4bit": load_in_4bit_enabled(),
"acoustic_hint_enabled": acoustic_hint_enabled(),
"audio_alignment_enabled": audio_alignment_enabled(),
"zero_gpu_size": zero_gpu_size(),
"input_summary": LAST_INPUT_SUMMARY,
"traceback": traceback.format_exc(limit=8),
}
return f"ZeroGPU inference failed: {type(exc).__name__}: {exc}", json.dumps(payload, indent=2)
def analyze(audio: str | tuple[int, np.ndarray] | None, target_text: str) -> tuple[str, str]:
return analyze_with_errors(audio, target_text)
@spaces.GPU(duration=5, size=zero_gpu_size())
def zero_gpu_healthcheck() -> str:
return "ok"
def analysis_started(browser_recording_payload: str, uploaded_audio: str | tuple[int, np.ndarray] | None) -> tuple[str, str]:
if not browser_recording_payload.strip() and uploaded_audio is None:
return (
"No clip ready yet. Use the browser recorder or upload a short speech clip first.",
json.dumps({"status": "waiting_for_audio"}, indent=2),
)
return (
"Checking the recording quality and acoustic evidence...",
json.dumps({"status": "running", "stage": "audio_preflight_then_acoustic_gate"}, indent=2),
)
def analyze_ui(
browser_recording_payload: str,
uploaded_audio: str | tuple[int, np.ndarray] | None,
target_text: str,
) -> tuple[str, str]:
selected_audio: str | tuple[int, np.ndarray] | None = (
browser_recording_payload.strip() if browser_recording_payload.strip() else uploaded_audio
)
if selected_audio is None:
return (
"No clip ready yet. Use the browser recorder or upload a short speech clip first.",
json.dumps({"status": "waiting_for_audio"}, indent=2),
)
return analyze_with_errors(selected_audio, target_text)
def build_app() -> gr.Blocks:
with gr.Blocks(title="Lisper ZeroGPU", theme=gr.themes.Soft()) as demo:
gr.Markdown(
"""
# Lisper ZeroGPU
Server-side Gemma 4 audio analysis for users whose browser cannot comfortably run the WebGPU model.
The currently validated fine-tuned Lisper model is Gemma 4 E2B. E4B and 31B are future model targets and should be deployed as separate revisions after training/eval.
"""
)
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("### Browser recorder")
browser_recording_payload = gr.Textbox(
label="Browser recorder payload",
visible=False,
elem_id="lisper-browser-recorder-payload",
)
browser_recording_status = gr.Markdown(
"No browser recording ready. This path bypasses Gradio's microphone recorder."
)
browser_recording_playback = gr.HTML("")
with gr.Row():
browser_record = gr.Button("Record", variant="primary")
browser_stop = gr.Button("Stop")
browser_clear = gr.Button("Clear")
gr.Markdown("### Upload fallback")
audio = gr.Audio(
sources=["upload"],
type="filepath",
label="Speech clip upload",
editable=False,
waveform_options=gr.WaveformOptions(show_recording_waveform=False),
)
audio_status = gr.Markdown(
"No uploaded clip ready. Use the browser recorder above, or upload an audio file here."
)
target_text = gr.Textbox(
label="Expected text",
placeholder="Example: Sally sells seashells.",
lines=2,
)
run = gr.Button("Analyze", variant="primary")
with gr.Column(scale=1):
output = gr.Textbox(label="Gemma response", lines=8)
parsed = gr.Code(label="Parsed JSON", language="json")
gr.Markdown(
f"""
**Configured model:** `{model_id()}`
**Configured adapter:** `{adapter_id() or "none"}`
**Adapter 4-bit load:** `{load_in_4bit_enabled()}`
**Acoustic hint:** `{acoustic_hint_enabled()}`
**Audio token alignment:** `{audio_alignment_enabled()}`
**ZeroGPU size:** `{zero_gpu_size()}`
If this Space errors on private or gated models, add `HF_TOKEN` as a Space secret. For local development without downloading the model, set `LISPER_ZERO_GPU_EAGER_LOAD=0`.
"""
)
browser_record.click(
None,
inputs=[browser_recording_payload],
outputs=[browser_recording_payload, browser_recording_status, browser_recording_playback],
js=BROWSER_RECORDER_START_JS,
queue=False,
show_progress="hidden",
)
browser_stop.click(
None,
inputs=[],
outputs=[browser_recording_payload, browser_recording_status, browser_recording_playback],
js=BROWSER_RECORDER_STOP_JS,
queue=False,
show_progress="hidden",
)
browser_clear.click(
None,
inputs=[],
outputs=[browser_recording_payload, browser_recording_status, browser_recording_playback],
js=BROWSER_RECORDER_CLEAR_JS,
queue=False,
show_progress="hidden",
)
audio.change(
lambda: "Uploaded clip ready. Analyze is available.",
inputs=[],
outputs=[audio_status],
queue=False,
show_progress="hidden",
)
audio.clear(
lambda: "No uploaded clip ready. Use the browser recorder above, or upload an audio file here.",
inputs=[],
outputs=[audio_status],
queue=False,
show_progress="hidden",
)
run.click(
analysis_started,
inputs=[browser_recording_payload, audio],
outputs=[output, parsed],
queue=False,
).then(
analyze_ui,
inputs=[browser_recording_payload, audio, target_text],
outputs=[output, parsed],
api_name="analyze",
)
return demo
demo = build_app()
if __name__ == "__main__":
demo.queue(default_concurrency_limit=1).launch(show_error=True)
|