Voice prompt md5 two tier cache

Voice prompt MD5 two-tier cache

When to use

Voice-cloning TTS builds a "voice prompt" (tokenized reference audio + embedding tensors) from a user's reference clip. The prompt is deterministic per (audio_bytes, reference_text) and takes hundreds of ms to rebuild. You want repeat generations to skip that cost entirely.

Steps

1. Compute the key as md5(audio_bytes || reference_text.encode("utf-8")). MD5 is fine here — you're not defending against a collision attacker, you're deduping content.
2. First tier: a process-local dict[str, VoicePrompt] keyed by the hex digest. O(1) on a hit and skips disk I/O entirely.
3. Second tier: <data_dir>/cache/<key>.prompt serialized via torch.save. torch.save handles both raw tensors and dict-of-tensors, so the same file format survives a schema change from tensor-only to dict-of-tensors.
4. On load, use torch.load(path, weights_only=True). It's safer against malicious pickles if you ever need to import a cache file from elsewhere.
5. On a corrupted cache file (any exception during load), unlink it and fall through to recomputation. Never raise — the cache must be best-effort.
6. Write-through: on cache miss, compute, populate both tiers. Expose a cache-clear that wipes both tiers and any derived artefacts (e.g. combined audio WAVs) under the same directory.
7. Provide clear_profile_cache(profile_id) that removes only combined_<profile_id>_*.wav etc. — useful when a profile is edited without wiping the entire cache.

Counter / Caveats

• Include the reference text in the hash. Two clones with the same audio but different reference transcripts produce different prompts and must not collide.
• Do not pickle the tensors with pickle directly — torch.save handles the device and dtype metadata across versions.
• Bound the RAM tier if you ever accept many distinct profiles per session; dict grows without limit. An OrderedDict + move_to_end LRU is a cheap upgrade.
• If users move the app's data_dir between platforms, treat the cache as ephemeral — .prompt files are torch-version sensitive and may not load on a newer torch build. Add a CACHE_VERSION prefix to the filename if you want eager invalidation.

Source references: backend/utils/cache.py.