Fix slow build: strip runtime-irrelevant heavy packages from requirements.txt

Removed from requirements.txt (causes slow builds, unused at Space runtime):
- torch / torchaudio pins → Spaces pre-installs correct CUDA build; re-pinning
forces pip to re-download ~1 GB on every build
- f5-tts>=1.0.0 → ~2 GB install with vocos/x-transformers/jieba/einops;
src/tts/f5_tts.py already falls back to MMS-TTS on ImportError
- speechbrain → heavy ML package, speaker-id not on critical path
- peft → training-only (Kaggle notebook), removed from app pipeline
- evaluate → training-only
- bitsandbytes → quantization library, not usable on cpu-basic anyway
- audiomentations → training-time augmentation only
- pytest / pytest-asyncio → dev-only, never run in production Space

Estimated build time improvement: 10-15 min -> ~2 min

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

requirements.txt

@@ -1,34 +1,30 @@
 # -----------------------------------------------------------------------------
-# Sahel-Agri Voice AI — Python Dependencies
-# HuggingFace Spaces (ZeroGPU) deployment — CUDA pre-installed, no +cu128 suffix
+# Sahel-Agri Voice AI — HuggingFace Spaces (cpu-basic) runtime dependencies
 #
-# Local CPU test:
-#   pip install -r requirements.txt
+# RULES:
+#   - Do NOT pin torch/torchaudio — Spaces pre-installs the correct CUDA build.
+#     Re-pinning them forces a 1 GB re-download on every build.
+#   - Do NOT add training-only packages (peft, evaluate, bitsandbytes,
+#     audiomentations) — they are only needed in the Kaggle notebook.
+#   - Do NOT add dev tools (pytest, pytest-asyncio) — not run in production.
+#   - f5-tts is intentionally omitted — it is a ~2 GB install and is only
+#     needed for GPU voice cloning.  src/tts/f5_tts.py falls back to MMS-TTS
+#     gracefully when the package is absent.
+#   - speechbrain is intentionally omitted — it is heavy and speaker-id is
+#     not on the critical path at runtime.
 # -----------------------------------------------------------------------------
 
-# PyTorch (CPU build — works on HF Spaces cpu-basic and locally)
-torch==2.11.0
-torchaudio==2.11.0
-
 # HuggingFace core
 transformers==5.5.0
 datasets==4.8.4
 accelerate==1.13.0
-evaluate==0.4.2
 huggingface-hub==1.9.0
 
-# PEFT (LoRA adapters)
-peft==0.18.1
-
 # Audio processing
 librosa==0.10.2
 soundfile==0.12.1
-audiomentations==0.43.1
-
-# Quantization (CPU: installs fine; 4-bit/8-bit requires GPU at runtime)
-bitsandbytes==0.49.2
 
-# Metrics
+# Metrics (used in app.py CER display)
 jiwer==3.0.4
 
 # Config & environment
@@ -41,33 +37,10 @@ python-dotenv==1.1.0
 # Pydantic v2
 pydantic==2.11.3
 
-# Testing
-pytest==8.3.5
-pytest-asyncio==0.26.0
-
 # Utilities
 numpy==2.2.4
 scipy==1.15.2
-
-# Phrase matching (fuzzy match for Whisper mis-transcriptions of Bambara/Fula)
 rapidfuzz==3.13.0
 
-# Voice cloning — F5-TTS (flow-matching, language-agnostic, reference-speaker)
-# Requires GPU at runtime (~750 MB model auto-downloaded on first use).
-# Falls back to MMS-TTS gracefully when not installed or GPU unavailable.
-f5-tts>=1.0.0
-
 # Kaggle API (used by Self-Teaching tab to trigger training runs)
 kaggle>=1.6.0
-
-# Speaker identification (ECAPA-TDNN 192-d embeddings, used by SpeakerProfileManager)
-speechbrain>=0.5.15
-
-# Voice cloning (OpenVoice V2 / myshell-openvoice) is NOT listed here.
-# myshell-openvoice pins librosa==0.9.1 which conflicts with our librosa==0.10.2.
-# VoiceCloner in src/tts/voice_cloner.py degrades gracefully (returns None) when
-# the package is absent — base VITS voice is used instead.
-# Re-enable once OpenVoice releases a version compatible with librosa>=0.10.
-
-# maliba-ai is NOT listed here — it has strict conflicting pins (librosa, soundfile).
-# It is installed lazily at runtime on first Bambara TTS call (see src/tts/waxal_tts.py).