diff --git a/Makefile b/Makefile
index 2267eeb..df58942 100644
--- a/Makefile
+++ b/Makefile
@@ -30,7 +30,7 @@ test: ## Тесты бэкенда
 typecheck: ## Проверить фронтенд компилятором
 	npm --prefix frontend run typecheck
 
-models: ## Скачать модели в backend/models/ (сейчас — эмбеддинги; GigaAM и Silero — lct-02)
+models: ## Скачать модели в backend/models/: эмбеддинги, GigaAM, Silero VAD; Silero TTS — вручную
 	cd backend && $(UV) run python scripts/models.py
 
 seed: ## Залить сценарии из /scenarios в БД
@@ -39,6 +39,9 @@ seed: ## Залить сценарии из /scenarios в БД
 lesson: ## Запустить занятие и напечатать ссылки: make lesson s=<сценарий> m=<режим>
 	cd backend && $(UV) run --no-project --with websockets python scripts/start_lesson.py "$(s)" "$(m)"
 
+latency: ## Замер задержки голосового контура по этапам — запускать на демо-машине
+	cd backend && $(UV) run --extra voice python scripts/latency.py
+
 migrate: ## Накатить миграции
 	cd backend && $(UV) run alembic upgrade head
 
@@ -54,4 +57,4 @@ pregen: ## Дерево диалога и WAV первых реплик для 
 demo: ## Поднять всё в демо-режиме: офлайн, прогретые модели
 	@echo "не реализовано — карточка tasks/lct-21-demo-readiness.md"; exit 1
 
-.PHONY: help dev down back front types test typecheck lesson migrate revision models seed repl pregen demo
+.PHONY: help dev down back front types test typecheck lesson latency migrate revision models seed repl pregen demo
diff --git a/backend/pyproject.toml b/backend/pyproject.toml
index c2cc502..f456c47 100644
--- a/backend/pyproject.toml
+++ b/backend/pyproject.toml
@@ -26,12 +26,24 @@ dependencies = [
 voice = [
     "onnx-asr>=0.6",
     "torch>=2.2",
+    # Нужен пакету Silero TTS: импорт зашит внутри v5_ru.pt.
+    "scipy>=1.11",
 ]
 dev = [
     "pytest>=8.2",
     "pytest-asyncio>=0.23",
 ]
 
+# torch — только CPU-сборка. Из обычного PyPI он тянет CUDA-библиотеки
+# на несколько гигабайт, а GPU на стенде нет (docs/arch/STACK.md).
+[tool.uv.sources]
+torch = { index = "pytorch-cpu" }
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
diff --git a/backend/scripts/latency.py b/backend/scripts/latency.py
new file mode 100644
index 0000000..41dc351
--- /dev/null
+++ b/backend/scripts/latency.py
@@ -0,0 +1,222 @@
+"""make latency: замер бюджета задержки голосового контура по этапам.
+
+Запускать **на той машине, которая поедет на занятие** (docs/arch/BACKEND.md).
+Цифры пишутся в docs/LATENCY.md руками вместе с решениями — скрипт только меряет.
+
+Что меряется и что нет:
+  * STT — GigaAM v3 RNNT и CTC (int8) на живой речи Golos: задержка и ошибка
+    распознавания (WER) против эталонной расшифровки;
+  * VAD — Silero VAD на одном окне;
+  * TTS — Silero v5 на коротких репликах паникующего звонящего;
+  * эмбеддинги — multilingual-e5-small на реплике оператора.
+  * LLM не меряется: нужен ключ и сеть до провайдера.
+
+Golos — чистая речь со смартфонов, читающих фразы. Паникующий звонящий и оператор
+в стрессе дадут ошибку выше: цифра WER отсюда — нижняя граница.
+"""
+
+import json
+import os
+import platform
+import statistics
+import sys
+import time
+import wave
+from pathlib import Path
+
+import numpy as np
+
+ROOT = Path(__file__).resolve().parents[1]
+MODELS = ROOT / "models"
+sys.path.insert(0, str(ROOT))
+
+WARMUP = 2
+
+
+def machine() -> str:
+    cpu = "?"
+    try:
+        for line in open("/proc/cpuinfo", encoding="utf-8"):
+            if line.startswith("model name"):
+                cpu = line.split(":", 1)[1].strip()
+                break
+    except OSError:
+        cpu = platform.processor()
+    ram = os.sysconf("SC_PAGE_SIZE") * os.sysconf("SC_PHYS_PAGES") / 2**30
+    wsl = "WSL2" if "microsoft" in platform.release().lower() else "нативно"
+    return f"{cpu}, ядер {os.cpu_count()}, RAM {ram:.0f} ГБ, {platform.system()} {wsl}"
+
+
+def stats(values_ms: list[float]) -> str:
+    ordered = sorted(values_ms)
+    p95 = ordered[min(len(ordered) - 1, int(round(0.95 * (len(ordered) - 1))))]
+    digits = 2 if p95 < 10 else 0  # VAD укладывается в доли миллисекунды
+    return f"медиана {statistics.median(ordered):.{digits}f} мс, p95 {p95:.{digits}f} мс"
+
+
+def read_wav(path: Path) -> tuple[np.ndarray, int]:
+    with wave.open(str(path)) as w:
+        rate, channels, width = w.getframerate(), w.getnchannels(), w.getsampwidth()
+        raw = w.readframes(w.getnframes())
+    assert width == 2, f"{path.name}: ожидался PCM16"
+    audio = np.frombuffer(raw, dtype=np.int16).astype(np.float32) / 32768
+    if channels > 1:
+        audio = audio.reshape(-1, channels).mean(axis=1)
+    return audio, rate
+
+
+def normalize(text: str) -> list[str]:
+    text = text.lower().replace("ё", "е")
+    return "".join(ch if ch.isalnum() or ch.isspace() else " " for ch in text).split()
+
+
+def wer(reference: list[str], hypothesis: list[str]) -> tuple[int, int]:
+    """Расстояние Левенштейна по словам. Возвращает (ошибок, слов в эталоне)."""
+    prev = list(range(len(hypothesis) + 1))
+    for i, ref_word in enumerate(reference, 1):
+        cur = [i] + [0] * len(hypothesis)
+        for j, hyp_word in enumerate(hypothesis, 1):
+            cur[j] = min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ref_word != hyp_word))
+        prev = cur
+    return prev[-1], len(reference)
+
+
+def measure_stt() -> None:
+    import onnx_asr
+
+    corpus = MODELS / "samples" / "golos"
+    manifest = json.load(open(corpus / "manifest.json", encoding="utf-8-sig"))["items"]
+    clips = [(read_wav(corpus / item["file"]), item["text"]) for item in manifest]
+    total_audio = sum(len(audio) / rate for (audio, rate), _ in clips)
+    print(f"\n## STT — Golos, {len(clips)} фраз, {total_audio:.0f} с речи")
+
+    for name in ("gigaam-v3-rnnt", "gigaam-v3-ctc"):
+        started = time.monotonic()
+        model = onnx_asr.load_model(name, MODELS / "gigaam-v3-onnx", quantization="int8")
+        load_s = time.monotonic() - started
+
+        for (audio, rate), _ in clips[:WARMUP]:
+            model.recognize(audio, sample_rate=rate)
+
+        latencies, errors, words, audio_s, compute_s = [], 0, 0, 0.0, 0.0
+        worst, durations = [], []
+        for (audio, rate), reference in clips:
+            started = time.monotonic()
+            hypothesis = model.recognize(audio, sample_rate=rate)
+            elapsed = time.monotonic() - started
+            latencies.append(elapsed * 1000)
+            durations.append(len(audio) / rate)
+            audio_s += len(audio) / rate
+            compute_s += elapsed
+            e, n = wer(normalize(reference), normalize(hypothesis))
+            errors, words = errors + e, words + n
+            if e:
+                worst.append(f"    «{reference}» → «{hypothesis}»")
+
+        print(f"- {name} int8: загрузка {load_s:.1f} с; фраза {stats(latencies)}; "
+              f"RTF {compute_s / audio_s:.3f}; WER {100 * errors / words:.1f}% ({errors} из {words} слов)")
+        # Время растёт с длиной фразы, а у Golos фразы длиннее вопросов оператора:
+        # медиана по корпусу завышает задержку. Считаем зависимость от длительности.
+        slope, intercept = np.polyfit(durations, latencies, 1)
+        span = f"{min(durations):.1f}–{max(durations):.1f} с"
+        print(f"  зависимость: ≈ {intercept:.0f} мс + {slope:.0f} мс × секунд речи (фразы корпуса {span})")
+        print("  " + "; ".join(f"{sec} с → ~{intercept + slope * sec:.0f} мс" for sec in (1.5, 2, 3)))
+        for line in worst[:5]:
+            print(line)
+
+
+def measure_vad() -> None:
+    import onnxruntime as ort
+
+    path = MODELS / "silero-vad" / "silero_vad.onnx"
+    session = ort.InferenceSession(str(path), providers=["CPUExecutionProvider"])
+    inputs = {i.name: i for i in session.get_inputs()}
+    window = 512  # окно Silero VAD при 16 кГц — 32 мс
+    feed = {"input": np.zeros((1, window), dtype=np.float32), "sr": np.array(16000, dtype=np.int64)}
+    if "state" in inputs:
+        feed["state"] = np.zeros((2, 1, 128), dtype=np.float32)
+    else:
+        feed["h"] = np.zeros((2, 1, 64), dtype=np.float32)
+        feed["c"] = np.zeros((2, 1, 64), dtype=np.float32)
+    for _ in range(50):
+        session.run(None, feed)
+    runs = []
+    for _ in range(500):
+        started = time.monotonic()
+        session.run(None, feed)
+        runs.append((time.monotonic() - started) * 1000)
+    print(f"\n## VAD — Silero, окно 32 мс\n- на окно: {stats(runs)} (задержка endpointing задана конфигом: 600 мс)")
+
+
+def measure_tts() -> None:
+    path = MODELS / "silero-tts" / "v5_ru.pt"
+    if not path.exists():
+        print("\n## TTS\n- не измерено: нет models/silero-tts/v5_ru.pt")
+        return
+    import torch
+
+    torch.set_num_threads(4)
+    # Профилирующий компилятор TorchScript на новых длинах входа ничего не выигрывает:
+    # без него синтез на 13% быстрее на тех же фразах.
+    torch._C._jit_set_profiling_executor(False)
+    started = time.monotonic()
+    importer = torch.package.PackageImporter(str(path))
+    model = importer.load_pickle("tts_models", "model")
+    model.to(torch.device("cpu"))
+    load_s = time.monotonic() - started
+    speakers = getattr(model, "speakers", [])
+    speaker = "xenia" if "xenia" in speakers else speakers[0]
+
+    warmup = ["Алло!", "Горит балкон на пятом этаже!", "Скорее приезжайте, пожалуйста, мы задыхаемся!"]
+    for text in warmup:
+        model.apply_tts(text=text, speaker=speaker, sample_rate=24000)
+
+    # Реплики паникующего звонящего разной длины: время синтеза растёт с длиной звука,
+    # и первая фраза звонящего — самая важная для ощущения «ответил сразу».
+    lines = [
+        "Алло! Помогите!",
+        "Горим!",
+        "Дым идёт в подъезд!",
+        "Жена с ребёнком в дальней комнате!",
+        "Я не знаю, где перекрыть газ!",
+        "Быстрее, пожалуйста, дышать нечем!",
+        "Муж пытался потушить, но не получилось!",
+        "Пятый этаж, подъезд второй!",
+    ]
+    latencies, audio_s, compute_s = [], 0.0, 0.0
+    per_line = []
+    for text in lines:
+        started = time.monotonic()
+        audio = model.apply_tts(text=text, speaker=speaker, sample_rate=24000)
+        elapsed = time.monotonic() - started
+        latencies.append(elapsed * 1000)
+        audio_s += len(audio) / 24000
+        compute_s += elapsed
+        per_line.append(f"«{text}» {len(audio) / 24000:.1f} с звука → {elapsed * 1000:.0f} мс")
+    print(f"\n## TTS — Silero v5, голос {speaker}, 24 кГц, 4 потока, без профилирующего компилятора")
+    print(f"- загрузка {load_s:.1f} с; реплика {stats(latencies)}; RTF {compute_s / audio_s:.3f} "
+          f"(синтез в {audio_s / compute_s:.0f} раз быстрее реального времени)")
+    for line in per_line:
+        print(f"  {line}")
+    print(f"- голоса: {', '.join(speakers)}")
+
+
+def measure_embeddings() -> None:
+    from app.dialog.embeddings import E5Embedder
+
+    embedder = E5Embedder(MODELS / "e5-small")
+    embedder.embed(["разогрев"])
+    runs = []
+    for _ in range(50):
+        started = time.monotonic()
+        embedder.embed(["На каком этаже пожар?"])
+        runs.append((time.monotonic() - started) * 1000)
+    print(f"\n## Эмбеддинги — multilingual-e5-small int8\n- реплика: {stats(runs)}")
+
+
+if __name__ == "__main__":
+    print(f"# Замер задержки\n\nМашина: {machine()}")
+    only = set(sys.argv[1:])
+    for name, step in [("stt", measure_stt), ("vad", measure_vad), ("tts", measure_tts), ("emb", measure_embeddings)]:
+        if not only or name in only:
+            step()
diff --git a/backend/scripts/models.py b/backend/scripts/models.py
index 4fe5632..f7cbf3c 100644
--- a/backend/scripts/models.py
+++ b/backend/scripts/models.py
@@ -1,7 +1,8 @@
 """make models: веса моделей в backend/models/.
 
-Сейчас качает эмбеддинги для слот-автомата. Распознавание (GigaAM) и синтез
-(Silero) добавит карточка lct-02 — после замера на демо-машине.
+Эмбеддинги, распознавание (GigaAM v3, int8) и VAD качаются с Hugging Face.
+Синтез (Silero TTS v5) лежит на models.silero.ai — российском хосте, который
+не отвечает из-под VPN: его скрипт не качает, а проверяет и говорит, что делать.
 Докачка продолжается с места обрыва: сеть на стенде бывает медленной.
 """
 
@@ -13,12 +14,27 @@ ROOT = Path(__file__).resolve().parents[1]
 MODELS = ROOT / "models"
 
 E5 = "https://huggingface.co/Xenova/multilingual-e5-small/resolve/main/"
+GIGAAM = "https://huggingface.co/istupakov/gigaam-v3-onnx/resolve/main/"
+VAD = "https://huggingface.co/istupakov/silero-vad-onnx/resolve/main/"
 FILES = {
     "e5-small/config.json": E5 + "config.json",
     "e5-small/tokenizer.json": E5 + "tokenizer.json",
     "e5-small/model_quantized.onnx": E5 + "onnx/model_quantized.onnx",
+    # RNNT — основная модель, CTC — запасная на случай, если RNNT не загрузится
+    # (по скорости CTC не выигрывает — docs/LATENCY.md).
+    "gigaam-v3-onnx/config.json": GIGAAM + "config.json",
+    "gigaam-v3-onnx/v3_vocab.txt": GIGAAM + "v3_vocab.txt",
+    "gigaam-v3-onnx/v3_rnnt_encoder.int8.onnx": GIGAAM + "v3_rnnt_encoder.int8.onnx",
+    "gigaam-v3-onnx/v3_rnnt_decoder.int8.onnx": GIGAAM + "v3_rnnt_decoder.int8.onnx",
+    "gigaam-v3-onnx/v3_rnnt_joint.int8.onnx": GIGAAM + "v3_rnnt_joint.int8.onnx",
+    "gigaam-v3-onnx/v3_ctc.int8.onnx": GIGAAM + "v3_ctc.int8.onnx",
+    "silero-vad/config.json": VAD + "config.json",
+    "silero-vad/silero_vad.onnx": VAD + "silero_vad.onnx",
 }
 
+SILERO_TTS = "silero-tts/v5_ru.pt"
+SILERO_TTS_URL = "https://models.silero.ai/models/tts/ru/v5_ru.pt"
+
 
 def fetch(url: str, target: Path) -> None:
     partial = target.with_suffix(target.suffix + ".part")
@@ -43,7 +59,18 @@ def main() -> None:
             continue
         target.parent.mkdir(parents=True, exist_ok=True)
         fetch(url, target)
-    print("эмбеддинги готовы; GigaAM и Silero — карточка tasks/lct-02-latency-baseline.md")
+    tts = MODELS / SILERO_TTS
+    if tts.exists():
+        print(f"  {SILERO_TTS}: уже есть")
+        print("все модели на месте")
+        return 0
+    # Не качаем с зеркал: .pt грузится через pickle, файл из непроверенного
+    # источника — это чужой код на стенде.
+    print(f"""
+  {SILERO_TTS}: НЕТ. Хост models.silero.ai не отвечает из-под VPN — скачайте вручную:
+    {SILERO_TTS_URL}
+  и положите в backend/models/{SILERO_TTS}""")
+    return 1
 
 
 if __name__ == "__main__":
diff --git a/backend/uv.lock b/backend/uv.lock
index caf8d81..896a217 100644
--- a/backend/uv.lock
+++ b/backend/uv.lock
@@ -1,6 +1,10 @@
 version = 1
 revision = 3
 requires-python = "==3.11.*"
+resolution-markers = [
+    "sys_platform != 'darwin'",
+    "sys_platform == 'darwin'",
+]
 
 [[package]]
 name = "alembic"
@@ -90,75 +94,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" },
 ]
 
-[[package]]
-name = "cuda-bindings"
-version = "13.4.1"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "cuda-pathfinder" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/dc/45/57fb37fea9200d854960e36c2181675f92d16bfd7012cc5a95b153e782f0/cuda_bindings-13.4.1-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:56c1552e207b291c321cef9952fbfaf8591c2a14d9c6e5215071020f541a04e5", size = 6484208, upload-time = "2026-09-10T01:16:46.069Z" },
-    { url = "https://files.pythonhosted.org/packages/ff/00/15dc1b8f2e5e98975107c366ec5671256d32b8e5c5ed6c1bb3a90a64f035/cuda_bindings-13.4.1-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e5dc0f13cfd14cd62206fede462f91c497d0c484b8236f4158531d20377066cd", size = 7159324, upload-time = "2026-09-10T01:16:48.109Z" },
-]
-
-[[package]]
-name = "cuda-pathfinder"
-version = "1.8.1"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/9d/e6/22df83f82f9bc26cb1c42265cf14d34d4908dba2a0f261bd7b28244acb00/cuda_pathfinder-1.8.1-py3-none-any.whl", hash = "sha256:ae0137ff9e56ea97499bcbf54f5f2778ec25f3266715ac86da192a795af982a8", size = 62552, upload-time = "2026-09-02T16:55:28.64Z" },
-]
-
-[[package]]
-name = "cuda-toolkit"
-version = "13.0.3.0"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/d1/c7/a79086a62c98befcdb8349656c6f114e2db3b8b2422f6e25c97a7f2a9a3c/cuda_toolkit-13.0.3.0-py2.py3-none-any.whl", hash = "sha256:d693caaa261214ddd7dbb60d68e71cbed884e68c2be7509778f3051da0b91c3f", size = 2512, upload-time = "2026-04-14T00:50:08.173Z" },
-]
-
-[package.optional-dependencies]
-cublas = [
-    { name = "nvidia-cublas", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-cuda-nvrtc", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-cudart = [
-    { name = "nvidia-cuda-runtime", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-cufft = [
-    { name = "nvidia-cufft", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-cufile = [
-    { name = "nvidia-cufile", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux')" },
-]
-cupti = [
-    { name = "nvidia-cuda-cupti", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-curand = [
-    { name = "nvidia-curand", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-cusolver = [
-    { name = "nvidia-cublas", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-cusolver", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-cusparse", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-cusparse = [
-    { name = "nvidia-cusparse", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-    { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-nvjitlink = [
-    { name = "nvidia-nvjitlink", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-nvrtc = [
-    { name = "nvidia-cuda-nvrtc", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-nvtx = [
-    { name = "nvidia-nvtx", marker = "(platform_machine == 'aarch64' and sys_platform == 'linux') or (platform_machine == 'x86_64' and sys_platform == 'linux') or (platform_machine == 'AMD64' and sys_platform == 'win32')" },
-]
-
 [[package]]
 name = "fastapi"
 version = "0.141.1"
@@ -364,7 +299,9 @@ dev = [
 ]
 voice = [
     { name = "onnx-asr" },
-    { name = "torch" },
+    { name = "scipy" },
+    { name = "torch", version = "2.14.0", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform == 'darwin'" },
+    { name = "torch", version = "2.14.0+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "sys_platform != 'darwin'" },
 ]
 
 [package.metadata]
@@ -381,10 +318,11 @@ requires-dist = [
     { name = "pytest", marker = "extra == 'dev'", specifier = ">=8.2" },
     { name = "pytest-asyncio", marker = "extra == 'dev'", specifier = ">=0.23" },
     { name = "pyyaml", specifier = ">=6.0" },
+    { name = "scipy", marker = "extra == 'voice'", specifier = ">=1.11" },
     { name = "sqlalchemy", extras = ["asyncio"], specifier = ">=2.0" },
     { name = "structlog", specifier = ">=24.1" },
     { name = "tokenizers", specifier = ">=0.19" },
-    { name = "torch", marker = "extra == 'voice'", specifier = ">=2.2" },
+    { name = "torch", marker = "extra == 'voice'", specifier = ">=2.2", index = "https://download.pytorch.org/whl/cpu" },
     { name = "uvicorn", extras = ["standard"], specifier = ">=0.30" },
 ]
 provides-extras = ["voice", "dev"]
@@ -464,158 +402,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/15/ce/e5ec180bc41812edcd8daeb8639d205622c0e8c02259d8ab25a0201b3c2a/numpy-2.4.6-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:2803abfebfc990042cd494d8ce2d5f82e9d847af6d35ec486923aa19dbad5e73", size = 12504263, upload-time = "2026-05-18T23:37:09.715Z" },
 ]
 
-[[package]]
-name = "nvidia-cublas"
-version = "13.1.1.3"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-cuda-nvrtc" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/a7/a1/0bd24ee8c8d03adac032fd2909426a00c88f8c57961b1277ded97f91119f/nvidia_cublas-13.1.1.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:b7a210458267ac818974c53038fbec2e969d5c99f305ab15c72522fa9f001dd5", size = 542848918, upload-time = "2026-04-08T18:46:22.985Z" },
-    { url = "https://files.pythonhosted.org/packages/3b/cd/154ca20c38269e05eff77c1464e6c1da89f50a6390b565e9d82e06bc11e1/nvidia_cublas-13.1.1.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:37936a16db8fe4ac1f065c2139360608a543a09275cb1a1af612e08cfa065436", size = 423138758, upload-time = "2026-04-08T18:46:58.655Z" },
-]
-
-[[package]]
-name = "nvidia-cuda-cupti"
-version = "13.0.85"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/2a/2a/80353b103fc20ce05ef51e928daed4b6015db4aaa9162ed0997090fe2250/nvidia_cuda_cupti-13.0.85-py3-none-manylinux_2_25_aarch64.whl", hash = "sha256:796bd679890ee55fb14a94629b698b6db54bcfd833d391d5e94017dd9d7d3151", size = 10310827, upload-time = "2025-09-04T08:26:42.012Z" },
-    { url = "https://files.pythonhosted.org/packages/33/6d/737d164b4837a9bbd202f5ae3078975f0525a55730fe871d8ed4e3b952b0/nvidia_cuda_cupti-13.0.85-py3-none-manylinux_2_25_x86_64.whl", hash = "sha256:4eb01c08e859bf924d222250d2e8f8b8ff6d3db4721288cf35d14252a4d933c8", size = 10715597, upload-time = "2025-09-04T08:26:51.312Z" },
-]
-
-[[package]]
-name = "nvidia-cuda-nvrtc"
-version = "13.0.88"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/c3/68/483a78f5e8f31b08fb1bb671559968c0ca3a065ac7acabfc7cee55214fd6/nvidia_cuda_nvrtc-13.0.88-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:ad9b6d2ead2435f11cbb6868809d2adeeee302e9bb94bcf0539c7a40d80e8575", size = 90215200, upload-time = "2025-09-04T08:28:44.204Z" },
-    { url = "https://files.pythonhosted.org/packages/b7/dc/6bb80850e0b7edd6588d560758f17e0550893a1feaf436807d64d2da040f/nvidia_cuda_nvrtc-13.0.88-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d27f20a0ca67a4bb34268a5e951033496c5b74870b868bacd046b1b8e0c3267b", size = 43015449, upload-time = "2025-09-04T08:28:20.239Z" },
-]
-
-[[package]]
-name = "nvidia-cuda-runtime"
-version = "13.0.96"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/87/4f/17d7b9b8e285199c58ce28e31b5c5bbaa4d8271af06a89b6405258245de2/nvidia_cuda_runtime-13.0.96-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:ef9bcbe90493a2b9d810e43d249adb3d02e98dd30200d86607d8d02687c43f55", size = 2261060, upload-time = "2025-10-09T08:55:15.78Z" },
-    { url = "https://files.pythonhosted.org/packages/2e/24/d1558f3b68b1d26e706813b1d10aa1d785e4698c425af8db8edc3dced472/nvidia_cuda_runtime-13.0.96-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:7f82250d7782aa23b6cfe765ecc7db554bd3c2870c43f3d1821f1d18aebf0548", size = 2243632, upload-time = "2025-10-09T08:55:36.117Z" },
-]
-
-[[package]]
-name = "nvidia-cudnn-cu13"
-version = "9.24.0.43"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-cublas" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/ca/30/7c257e3d5cb4fecb147b93895c66e29c93f8e76d74b45bb418ff0587c4ec/nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:a6812a554a1ff0413e9c52b84c26c050380649ab9615f9c16bded368ce9f421f", size = 650976863, upload-time = "2026-07-02T16:23:39.248Z" },
-    { url = "https://files.pythonhosted.org/packages/5c/ba/791cffd048fe5b044e620df55267e3e95c0e6e07d50b41e377c03dfc910f/nvidia_cudnn_cu13-9.24.0.43-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:71f181cd810e90f9b6023b01186fe82d13d65f0ec098581ee201d39fad769e4b", size = 553099438, upload-time = "2026-07-02T16:27:42.58Z" },
-]
-
-[[package]]
-name = "nvidia-cufft"
-version = "12.0.0.61"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-nvjitlink" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/8b/ae/f417a75c0259e85c1d2f83ca4e960289a5f814ed0cea74d18c353d3e989d/nvidia_cufft-12.0.0.61-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2708c852ef8cd89d1d2068bdbece0aa188813a0c934db3779b9b1faa8442e5f5", size = 214053554, upload-time = "2025-09-04T08:31:38.196Z" },
-    { url = "https://files.pythonhosted.org/packages/a8/2f/7b57e29836ea8714f81e9898409196f47d772d5ddedddf1592eadb8ab743/nvidia_cufft-12.0.0.61-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:6c44f692dce8fd5ffd3e3df134b6cdb9c2f72d99cf40b62c32dde45eea9ddad3", size = 214085489, upload-time = "2025-09-04T08:31:56.044Z" },
-]
-
-[[package]]
-name = "nvidia-cufile"
-version = "1.15.1.6"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/3f/70/4f193de89a48b71714e74602ee14d04e4019ad36a5a9f20c425776e72cd6/nvidia_cufile-1.15.1.6-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:08a3ecefae5a01c7f5117351c64f17c7c62efa5fffdbe24fc7d298da19cd0b44", size = 1223672, upload-time = "2025-09-04T08:32:22.779Z" },
-    { url = "https://files.pythonhosted.org/packages/ab/73/cc4a14c9813a8a0d509417cf5f4bdaba76e924d58beb9864f5a7baceefbf/nvidia_cufile-1.15.1.6-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:bdc0deedc61f548bddf7733bdc216456c2fdb101d020e1ab4b88d232d5e2f6d1", size = 1136992, upload-time = "2025-09-04T08:32:14.119Z" },
-]
-
-[[package]]
-name = "nvidia-curand"
-version = "10.4.0.35"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/1e/72/7c2ae24fb6b63a32e6ae5d241cc65263ea18d08802aaae087d9f013335a2/nvidia_curand-10.4.0.35-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:133df5a7509c3e292aaa2b477afd0194f06ce4ea24d714d616ff36439cee349a", size = 61962106, upload-time = "2025-08-04T10:21:41.128Z" },
-    { url = "https://files.pythonhosted.org/packages/a5/9f/be0a41ca4a4917abf5cb9ae0daff1a6060cc5de950aec0396de9f3b52bc5/nvidia_curand-10.4.0.35-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:1aee33a5da6e1db083fe2b90082def8915f30f3248d5896bcec36a579d941bfc", size = 59544258, upload-time = "2025-08-04T10:22:03.992Z" },
-]
-
-[[package]]
-name = "nvidia-cusolver"
-version = "12.0.4.66"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-cublas" },
-    { name = "nvidia-cusparse" },
-    { name = "nvidia-nvjitlink" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/c8/c3/b30c9e935fc01e3da443ec0116ed1b2a009bb867f5324d3f2d7e533e776b/nvidia_cusolver-12.0.4.66-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:02c2457eaa9e39de20f880f4bd8820e6a1cfb9f9a34f820eb12a155aa5bc92d2", size = 223467760, upload-time = "2025-09-04T08:33:04.222Z" },
-    { url = "https://files.pythonhosted.org/packages/5f/67/cba3777620cdacb99102da4042883709c41c709f4b6323c10781a9c3aa34/nvidia_cusolver-12.0.4.66-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:0a759da5dea5c0ea10fd307de75cdeb59e7ea4fcb8add0924859b944babf1112", size = 200941980, upload-time = "2025-09-04T08:33:22.767Z" },
-]
-
-[[package]]
-name = "nvidia-cusparse"
-version = "12.6.3.3"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-nvjitlink" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/f8/94/5c26f33738ae35276672f12615a64bd008ed5be6d1ebcb23579285d960a9/nvidia_cusparse-12.6.3.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:80bcc4662f23f1054ee334a15c72b8940402975e0eab63178fc7e670aa59472c", size = 162155568, upload-time = "2025-09-04T08:33:42.864Z" },
-    { url = "https://files.pythonhosted.org/packages/fa/18/623c77619c31d62efd55302939756966f3ecc8d724a14dab2b75f1508850/nvidia_cusparse-12.6.3.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2b3c89c88d01ee0e477cb7f82ef60a11a4bcd57b6b87c33f789350b59759360b", size = 145942937, upload-time = "2025-09-04T08:33:58.029Z" },
-]
-
-[[package]]
-name = "nvidia-cusparselt-cu13"
-version = "0.8.1"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/46/e1/cdc1797eadf82d3a9a575a19b33fdc871a97edbec42c00b5b5e914f4aff4/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_aarch64.whl", hash = "sha256:4dca476c50bf4780d46cd0bfbd82e2bc10a08e4fef7950917ce8d7578d22a23f", size = 221051344, upload-time = "2025-09-05T18:49:51.289Z" },
-    { url = "https://files.pythonhosted.org/packages/34/7d/2661f2fb3ac4302f3a246f5fc030213ac60c1fe0bce84f9783dbd831dbb7/nvidia_cusparselt_cu13-0.8.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:786ce87568c303fadb5afcc7102d454cd3040d75f6f8626f5db460d1871f4dd0", size = 170148586, upload-time = "2025-09-05T18:50:50.248Z" },
-]
-
-[[package]]
-name = "nvidia-nccl-cu13"
-version = "2.30.7"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/d1/21/a73174c6157101bdf1ffc22b517f76ff0082613989dd9bc8f43e8034caac/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:ca786ffa5a647c75d4d1f5cc72a6c4f537947e2ba8823d7c8aaf768e7a7b9f77", size = 215983881, upload-time = "2026-06-09T03:23:15.633Z" },
-    { url = "https://files.pythonhosted.org/packages/3f/34/c500f90c7ae641b8e0f98965b36b8a7ac79cc8b296e8d251fe3eb592ee54/nvidia_nccl_cu13-2.30.7-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:cefa7fdb9710efd0f39c5f1be1d61ff6fc9a996c451265bd7fbdcf9455ed4b50", size = 215965170, upload-time = "2026-06-09T03:23:39.73Z" },
-]
-
-[[package]]
-name = "nvidia-nvjitlink"
-version = "13.4.52"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/7f/c1/091f198d7f87e31d67fa9680eec8f8e4c6f889881f729b759db36ff01612/nvidia_nvjitlink-13.4.52-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:90401db7e5a580a5067a468b3086e0b65f3b96ab3c42524afd741efc8a0e150a", size = 42452221, upload-time = "2026-09-09T18:01:32.895Z" },
-    { url = "https://files.pythonhosted.org/packages/7e/a9/225ff51e80de170be880cb88e992193bc8134a51059cc0a3952f967f62c5/nvidia_nvjitlink-13.4.52-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:3a589e3732140839349545efd0db67426523459b12e6952c3c73b6d55900f200", size = 40419746, upload-time = "2026-09-09T18:01:25.103Z" },
-]
-
-[[package]]
-name = "nvidia-nvshmem-cu13"
-version = "3.4.5"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/dc/0f/05cc9c720236dcd2db9c1ab97fff629e96821be2e63103569da0c9b72f19/nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:6dc2a197f38e5d0376ad52cd1a2a3617d3cdc150fd5966f4aee9bcebb1d68fe9", size = 60215947, upload-time = "2025-09-06T00:32:20.022Z" },
-    { url = "https://files.pythonhosted.org/packages/3c/35/a9bf80a609e74e3b000fef598933235c908fcefcef9026042b8e6dfde2a9/nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:290f0a2ee94c9f3687a02502f3b9299a9f9fe826e6d0287ee18482e78d495b80", size = 60412546, upload-time = "2025-09-06T00:32:41.564Z" },
-]
-
-[[package]]
-name = "nvidia-nvtx"
-version = "13.0.85"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/c2/f3/d86c845465a2723ad7e1e5c36dcd75ddb82898b3f53be47ebd429fb2fa5d/nvidia_nvtx-13.0.85-py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:4936d1d6780fbe68db454f5e72a42ff64d1fd6397df9f363ae786930fd5c1cd4", size = 148047, upload-time = "2025-09-04T08:29:01.761Z" },
-    { url = "https://files.pythonhosted.org/packages/a8/64/3708a90d1ebe202ffdeb7185f878a3c84d15c2b2c31858da2ce0583e2def/nvidia_nvtx-13.0.85-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:cb7780edb6b14107373c835bf8b72e7a178bac7367e23da7acb108f973f157a6", size = 148878, upload-time = "2025-09-04T08:28:53.627Z" },
-]
-
 [[package]]
 name = "onnx-asr"
 version = "0.12.0"
@@ -810,6 +596,27 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/da/e3/ea007450a105ae919a72393cb06f122f288ef60bba2dc64b26e2646fa315/pyyaml-6.0.3-cp311-cp311-win_amd64.whl", hash = "sha256:9f3bfb4965eb874431221a3ff3fdcddc7e74e3b07799e0e84ca4a0f867d449bf", size = 158763, upload-time = "2025-09-25T21:32:09.96Z" },
 ]
 
+[[package]]
+name = "scipy"
+version = "1.17.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/75/b4ce781849931fef6fd529afa6b63711d5a733065722d0c3e2724af9e40a/scipy-1.17.1-cp311-cp311-macosx_10_14_x86_64.whl", hash = "sha256:1f95b894f13729334fb990162e911c9e5dc1ab390c58aa6cbecb389c5b5e28ec", size = 31613675, upload-time = "2026-02-23T00:16:00.13Z" },
+    { url = "https://files.pythonhosted.org/packages/f7/58/bccc2861b305abdd1b8663d6130c0b3d7cc22e8d86663edbc8401bfd40d4/scipy-1.17.1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:e18f12c6b0bc5a592ed23d3f7b891f68fd7f8241d69b7883769eb5d5dfb52696", size = 28162057, upload-time = "2026-02-23T00:16:09.456Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/ee/18146b7757ed4976276b9c9819108adbc73c5aad636e5353e20746b73069/scipy-1.17.1-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:a3472cfbca0a54177d0faa68f697d8ba4c80bbdc19908c3465556d9f7efce9ee", size = 20334032, upload-time = "2026-02-23T00:16:17.358Z" },
+    { url = "https://files.pythonhosted.org/packages/ec/e6/cef1cf3557f0c54954198554a10016b6a03b2ec9e22a4e1df734936bd99c/scipy-1.17.1-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:766e0dc5a616d026a3a1cffa379af959671729083882f50307e18175797b3dfd", size = 22709533, upload-time = "2026-02-23T00:16:25.791Z" },
+    { url = "https://files.pythonhosted.org/packages/4d/60/8804678875fc59362b0fb759ab3ecce1f09c10a735680318ac30da8cd76b/scipy-1.17.1-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:744b2bf3640d907b79f3fd7874efe432d1cf171ee721243e350f55234b4cec4c", size = 33062057, upload-time = "2026-02-23T00:16:36.931Z" },
+    { url = "https://files.pythonhosted.org/packages/09/7d/af933f0f6e0767995b4e2d705a0665e454d1c19402aa7e895de3951ebb04/scipy-1.17.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:43af8d1f3bea642559019edfe64e9b11192a8978efbd1539d7bc2aaa23d92de4", size = 35349300, upload-time = "2026-02-23T00:16:49.108Z" },
+    { url = "https://files.pythonhosted.org/packages/b4/3d/7ccbbdcbb54c8fdc20d3b6930137c782a163fa626f0aef920349873421ba/scipy-1.17.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:cd96a1898c0a47be4520327e01f874acfd61fb48a9420f8aa9f6483412ffa444", size = 35127333, upload-time = "2026-02-23T00:17:01.293Z" },
+    { url = "https://files.pythonhosted.org/packages/e8/19/f926cb11c42b15ba08e3a71e376d816ac08614f769b4f47e06c3580c836a/scipy-1.17.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:4eb6c25dd62ee8d5edf68a8e1c171dd71c292fdae95d8aeb3dd7d7de4c364082", size = 37741314, upload-time = "2026-02-23T00:17:12.576Z" },
+    { url = "https://files.pythonhosted.org/packages/95/da/0d1df507cf574b3f224ccc3d45244c9a1d732c81dcb26b1e8a766ae271a8/scipy-1.17.1-cp311-cp311-win_amd64.whl", hash = "sha256:d30e57c72013c2a4fe441c2fcb8e77b14e152ad48b5464858e07e2ad9fbfceff", size = 36607512, upload-time = "2026-02-23T00:17:23.424Z" },
+    { url = "https://files.pythonhosted.org/packages/68/7f/bdd79ceaad24b671543ffe0ef61ed8e659440eb683b66f033454dcee90eb/scipy-1.17.1-cp311-cp311-win_arm64.whl", hash = "sha256:9ecb4efb1cd6e8c4afea0daa91a87fbddbce1b99d2895d151596716c0b2e859d", size = 24599248, upload-time = "2026-02-23T00:17:34.561Z" },
+]
+
 [[package]]
 name = "setuptools"
 version = "84.0.0"
@@ -908,28 +715,45 @@ wheels = [
 [[package]]
 name = "torch"
 version = "2.14.0"
-source = { registry = "https://pypi.org/simple" }
+source = { registry = "https://download.pytorch.org/whl/cpu" }
+resolution-markers = [
+    "sys_platform == 'darwin'",
+]
 dependencies = [
-    { name = "cuda-bindings", marker = "sys_platform == 'linux'" },
-    { name = "cuda-toolkit", extra = ["cublas", "cudart", "cufft", "cufile", "cupti", "curand", "cusolver", "cusparse", "nvjitlink", "nvrtc", "nvtx"], marker = "sys_platform == 'linux'" },
-    { name = "filelock" },
-    { name = "fsspec" },
-    { name = "jinja2" },
-    { name = "networkx" },
-    { name = "nvidia-cudnn-cu13", marker = "sys_platform == 'linux'" },
-    { name = "nvidia-cusparselt-cu13", marker = "sys_platform == 'linux'" },
-    { name = "nvidia-nccl-cu13", marker = "sys_platform == 'linux'" },
-    { name = "nvidia-nvshmem-cu13", marker = "sys_platform == 'linux'" },
-    { name = "setuptools" },
-    { name = "sympy" },
-    { name = "triton", marker = "sys_platform == 'linux'" },
-    { name = "typing-extensions" },
+    { name = "filelock", marker = "sys_platform == 'darwin'" },
+    { name = "fsspec", marker = "sys_platform == 'darwin'" },
+    { name = "jinja2", marker = "sys_platform == 'darwin'" },
+    { name = "networkx", marker = "sys_platform == 'darwin'" },
+    { name = "setuptools", marker = "sys_platform == 'darwin'" },
+    { name = "sympy", marker = "sys_platform == 'darwin'" },
+    { name = "typing-extensions", marker = "sys_platform == 'darwin'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/e6/2c/0c9757a1ea1fff8ccc4faff1d1c87e98a69542b5d4b606e50f7518f72e58/torch-2.14.0-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:ae530ddd3f3b94248b77f1fd3313c4f3405bd0d17283d505b81574816767133a", size = 127277244, upload-time = "2026-09-02T13:42:57.855Z" },
-    { url = "https://files.pythonhosted.org/packages/ed/30/eb2f6cf77cb8f2d5e7ce4343882a856783521f51b557673d78d1d7bb8661/torch-2.14.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:4dbea8a10d80b10dff2be1ed467d2617b704fe6bf76bbfad4194774c8f6d886b", size = 453993896, upload-time = "2026-09-02T13:43:29.231Z" },
-    { url = "https://files.pythonhosted.org/packages/8d/be/c0c8a845bf00552960d2f1f6a00b6ac9750fd312f55429c3e8c6c0e3aac4/torch-2.14.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:8d9e232b6376c62f3090237889fb1cb6887c0ece9f73e6c1052e80fc1481f999", size = 554583496, upload-time = "2026-09-02T13:44:10.199Z" },
-    { url = "https://files.pythonhosted.org/packages/e3/5b/ed9d177f6a133389d5fa89e298f3e2a879b62952aa928ac6ddd04467f225/torch-2.14.0-cp311-cp311-win_amd64.whl", hash = "sha256:6981872df75eb5409c439050d36b67cb57d88b4e11080516e2aa7410f6730705", size = 124096356, upload-time = "2026-09-02T13:43:02.941Z" },
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:1040f689e560ba6e02a0a147dfb9ec8dfdd3e4486ebec4b24dc8fdaa258fa9ba", upload-time = "2026-09-02T00:26:07Z" },
+]
+
+[[package]]
+name = "torch"
+version = "2.14.0+cpu"
+source = { registry = "https://download.pytorch.org/whl/cpu" }
+resolution-markers = [
+    "sys_platform != 'darwin'",
+]
+dependencies = [
+    { name = "filelock", marker = "sys_platform != 'darwin'" },
+    { name = "fsspec", marker = "sys_platform != 'darwin'" },
+    { name = "jinja2", marker = "sys_platform != 'darwin'" },
+    { name = "networkx", marker = "sys_platform != 'darwin'" },
+    { name = "setuptools", marker = "sys_platform != 'darwin'" },
+    { name = "sympy", marker = "sys_platform != 'darwin'" },
+    { name = "typing-extensions", marker = "sys_platform != 'darwin'" },
+]
+wheels = [
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0%2Bcpu-cp311-cp311-linux_s390x.whl", hash = "sha256:3fa39713d13f23563ad11636bfff76b677ab4addbc240c54e6b7e47622e8973d", upload-time = "2026-09-02T18:31:08Z" },
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0%2Bcpu-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:08894195f84541edcbd09072e6c53b791d4cdd09f650b05473f35718a848fd7b", upload-time = "2026-09-02T18:31:13Z" },
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0%2Bcpu-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:673dbf5c9bbadfffab7a386b6dd7a0c219f1408a328b7b4e86d0ae551cdafa42", upload-time = "2026-09-02T18:31:19Z" },
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0%2Bcpu-cp311-cp311-win_amd64.whl", hash = "sha256:8e2c47c6556c7d5a85848634372bb2252907d411e9cad669c99406856d536eb5", upload-time = "2026-09-02T18:31:24Z" },
+    { url = "https://download-r2.pytorch.org/whl/cpu/torch-2.14.0%2Bcpu-cp311-cp311-win_arm64.whl", hash = "sha256:38f0a0330267b36d76b988754878d94001c1519fe3743bd92f038ce5622bea5a", upload-time = "2026-09-02T18:31:28Z" },
 ]
 
 [[package]]
@@ -944,15 +768,6 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/a7/03/921a3d3c75785aca9ebfbfcabfbc3a1be12e2ab5265deb026d55a5a3f83e/tqdm-4.70.1-py3-none-any.whl", hash = "sha256:c293e525e6fef9c20e8728fd4612df02a0aa31bb5fe91ecd93e123b1b7bffa73", size = 80199, upload-time = "2026-09-11T07:25:14.599Z" },
 ]
 
-[[package]]
-name = "triton"
-version = "3.8.0"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/a3/cf/d21c9b1a4d1df9ba3aed218069f24f762a4b9565c5a362bef6f110c08e72/triton-3.8.0-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:372285307d4c44ee74cee32de0b4f04bd157e071427e38c6f4ee3e3beb2194f4", size = 226467015, upload-time = "2026-08-28T16:08:05.871Z" },
-    { url = "https://files.pythonhosted.org/packages/6f/9a/2c3d3823726d5ad5359f7065b02b984925158b74611c1e3987f6a37461fb/triton-3.8.0-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:68988ac85d5e7086baeda0ddc175af9667db7529b3c5e11a5c0601b8bef2200a", size = 247945226, upload-time = "2026-08-28T15:55:43.795Z" },
-]
-
 [[package]]
 name = "typing-extensions"
 version = "4.16.0"