TTS audio caches too broadly.

[jcc10]

LocalAI version:
localai/localai:v2.22.1-cublas-cuda12-ffmpeg@sha256:4ac028a056c946e047548e437952c80ba79b66af9428cbec628ab4ffedc47120
LocalAI Version v2.22.1 (015835dba2854572d50e167b7cade05af41ed214)

Environment, CPU architecture, OS, and Version:
While I am in docker, here is the bare meatal specs:
Linux pop-os 6.9.3-76060903-generic #202405300957~1726766035~22.04~4092a0e SMP PREEMPT_DYNAMIC Thu S x86_64 x86_64 x86_64 GNU/Linux

Describe the bug
Backend of TTS is not selecting properly, is caching too long, or is otherwise not re-generating when it should.

To Reproduce
Run the following commands:

curl --request POST \
  --url LocalAI_Instance/tts \
  --header 'content-type: application/json' \
  --data '{
  "backend": "coqui",
  "model": "tts_models/en/ljspeech/glow-tts",
  "input":"Welcome back my friends to the show that never ends!"
  }'|sha1sum

and

curl --request POST \
  --url  LocalAI_Instance/tts \
  --header 'content-type: application/json' \
  --data '{
   "backend": "vall-e-x",
   "input":"Welcome back my friends to the show that never ends!"
 }'|sha1sum

the two sums will be the same. Since they came from two different backends they shouldn't be.

Expected behavior
I would expect that two different backends should not produce identical results when given multiple arbitrary inputs.

Logs

5:51AM INF Success ip=127.0.0.1 latency="39.134µs" method=GET status=200 url=/readyz

5:51AM DBG guessDefaultsFromFile: modelPath is empty

5:51AM DBG Request for model: tts_models/en/ljspeech/glow-tts

5:51AM INF Loading model '' with backend coqui

5:51AM DBG Model already loaded in memory: 

5:51AM DBG Checking model availability ()

5:51AM INF Success ip=172.29.0.1 latency=169.709648ms method=POST status=200 url=/tts

5:52AM DBG guessDefaultsFromFile: modelPath is empty

5:52AM DBG Request for model: 

5:52AM INF Loading model '' with backend vall-e-x

5:52AM DBG Model already loaded in memory: 

5:52AM DBG Checking model availability ()

5:52AM INF Success ip=172.29.0.1 latency=70.942259ms method=POST status=200 url=/tts

Additional context
I intended to use the voice cloning, and the models trigger the correct backends to be prompted, however those files are also identical to the others.