docs: clarify openai-whisper build-isolation fix and .npy requirement

[moduvoice]

Problem

Two separate rough edges found while following the README verbatim on a fresh environment (Tesla T4 16GB, driver 550.163.01 / CUDA 12.4, Python 3.10.12):

1. pip install -r requirements.txt fails to build openai-whisper==20240930:

error: subprocess-exited-with-error
× Getting requirements to build wheel did not run successfully.
  ...
    File "<string>", line 5, in <module>
ModuleNotFoundError: No module named 'pkg_resources'
[end of output]
ERROR: Failed to build 'openai-whisper' when getting requirements to build wheel

This is a modern-pip/PEP517 build-isolation issue (the isolated build env doesn't include pkg_resources), not GPU/CUDA related — reproducible with a fresh venv + recent pip.

2. Using your own .wav (without a matching pre-extracted .npy) fails with a generic, unexplained error:

FileNotFoundError: [Errno 2] No such file or directory: '/path/to/my_own_voice.npy'

The README's "Model Download" section does explain that the WaveVAE encoder isn't released for security reasons and custom voices need a pre-extracted .npy requested via the linked Google Drive form — but that context is easy to miss, and the actual runtime error gives no hint that this is the reason. Someone naturally trying --input_wav my_own_recording.wav (a very natural first thing to try) hits a confusing crash with no clue why.

Fix

• README — add a note next to the pip install -r requirements.txt step: if the openai-whisper build fails with a pkg_resources error, retry with pip install --no-build-isolation -r requirements.txt (after installing torch/torchaudio first to keep the pinned versions).
• README — add a note near "Model Download" clarifying that --input_wav requires a matching .npy in the same directory, and that trying an arbitrary .wav without one will raise FileNotFoundError for the .npy — this is expected, not a bug, and is why the encoder-latent request process exists.

No code/behavior change — documentation only.

Verification (Tesla T4 16GB)

Working install (fixes issue 1):

python3 -m venv venv && source venv/bin/activate
pip install torch==2.6.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install --no-build-isolation -r requirements.txt

Reproduced the documented CLI example twice using the repo's own bundled assets/Chinese_prompt.wav (+ its matching .npy):

PYTHONPATH=. CUDA_VISIBLE_DEVICES=0 python tts/infer_cli.py \
  --input_wav 'assets/Chinese_prompt.wav' --input_text "这是一个测试。" --output_dir ./gen

Metric	Run 1	Run 2
Wall time (incl. model load)	66.4s	40.6s

Peak VRAM: 7.995 GB (fits Tesla T4's 16GB comfortably).

Also confirmed (not part of this PR, just verification notes): precision=torch.float16 is already correctly hardcoded for T4-class GPUs (tts/infer_cli.py:82), and all scaled_dot_product_attention call sites (tts/modules/ar_dur/commons/rot_transformer.py, tts/modules/aligner/whisper_small.py, tts/modules/llm_dit/transformer.py) don't force a specific attention backend, so they fall back correctly on Turing-class GPUs where the flash-attention kernel isn't available. No changes needed there.

Environment: Tesla T4 16GB · driver 550.163.01 (CUDA 12.4) · Python 3.10.12 · torch 2.6.0+cu124.

Raw metrics JSON

{"run1_wall_s": 66.406, "run2_wall_s": 40.594, "peak_vram_mib": 7995,
 "gpu": "Tesla T4", "driver": "550.163.01", "torch": "2.6.0+cu124"}