requirements.txt

Rama main · réplica de lectura del repositorio original.

# Core dependencies
setuptools>=65.0.0
wheel

# Audio dependencies (capture is ffmpeg-based; no PyAudio/portaudio needed)
SpeechRecognition>=3.10.0

# PyTorch - standard PyPI build supports CUDA (Linux) and MPS (macOS ARM) automatically.
# On Windows/Linux with NVIDIA GPU the installers add the CUDA wheel index:
#   pip install torch torchaudio --extra-index-url https://download.pytorch.org/whl/cu128
# Pinned: torch/ctranslate2/cuDNN must move together (ctranslate2 >= 4.5 needs
# the cuDNN 9 that torch 2.8 cu12x wheels bundle) — bump the trio deliberately.
torch==2.8.0
torchaudio==2.8.0

# Core ML dependencies
numpy<2.0.0
# The reference implementation, from PyPI — not the GitHub URL it used to be:
# uv shells out to the git CLI for VCS requirements, which stranded setup on
# machines without git (headless Macs with no Command Line Tools and no brew).
openai-whisper==20250625
faster-whisper==1.2.1
ctranslate2==4.7.1
transformers>=4.35.0
huggingface_hub>=0.19.0
accelerate>=0.24.0
datasets>=2.14.0

# Audio processing
soundfile>=0.12.0
# Deliberately unpinned, and deliberately not removed: panns-inference and
# torchlibrosa both require it, so a pin here would only fight them. It drags in
# the compiled scientific stack (numba/llvmlite, scikit-learn, scipy, soxr),
# which is where every source-build failure on a user machine has come from —
# see stt/wheel_policy.py, whose --only-binary list the installers pass so uv
# backtracks to a release with a wheel instead of trying to compile one.
librosa>=0.10.0
pydub>=0.25.1

# Audio tagging for music/speech detection (PANNs CNN14); pulls torchlibrosa
panns-inference>=0.1.0

# Web framework
flask>=2.3.0
flask-socketio>=5.3.0
flask-cors>=4.0.0
python-socketio>=5.9.0

# Text-to-Speech
edge-tts>=6.1.0
piper-tts>=1.2.0
# Local multilingual TTS: one ~385MB ONNX model covers 30 languages, so a
# language switch needs no per-language download the way piper does. Pulls
# onnxruntime (already required by piper-tts), numpy, soundfile and
# huggingface-hub — all of which we install anyway.
supertonic>=1.3.1

# Utilities
pytz>=2023.3
# IANA timezone database for stdlib zoneinfo. Required on Windows, which ships no
# system tz database — without it a configured timezone.value silently falls back to
# the machine's own zone. Pure data, no code.
tzdata>=2024.1
silero-vad>=4.0.0
smbprotocol>=1.10.0
# Live CPU/RAM metrics for the health dashboard (optional — degrades to "n/a" if absent)
psutil>=5.9.0
# Optional error reporting — only active when crash_reporting.sentry_dsn is set
sentry-sdk[flask]>=2.0,<3

# Optional: For better performance
# ffmpeg-python>=0.2.0  # Requires ffmpeg installed on system

# Optional: LLM translation with live_translation.llm.provider = "local".
# Runs a GGUF in-process — no separate server, installer or port — and unlike an NMT
# model it can be told the domain ("render biblical terminology as English Bibles
# do"), which fixes a class of error no decoding parameter reaches. Left commented out
# because it is a large build and most installs will not use it; the "endpoint"
# provider needs nothing extra and talks to any OpenAI-compatible server.
#
#   pip install llama-cpp-python
#
# The PyPI wheel is CPU-only on every platform, which works and is the safe default
# (a caption is short — measured p50 19 output tokens). For acceleration:
#   macOS ARM  — the PyPI wheel already includes Metal
#   CUDA       — needs a matching prebuilt wheel, the same pattern as torch above:
#                pip install llama-cpp-python \
#                  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/<cuda>
# llama-cpp-python>=0.3.0