https://mystic.ai is slicing big, not biggest, GPUs which reduces start up times, and is much cheaper, plus has inexpensive GPUs about half the price of lightning.ai. There is no coding environment though, just a place to run serverless pipelines that you create locally and push to their server. Or, you can use and pay for public pipelines on their servers. You literally pay by the sec, and on a share GPU you never have load time, so it is very cost effective, although you may have to wait in a queue. I haven't fully wrapped my head around creating and loading my own pipelines, but there are a number of public models that you can just use, including the ASR(TTS)/STT that I am looking for.
Free $20, credit to figure out how they work. I have found the sec billing to mean that it is quite cheap to explore even with GPUs compared to the 10 min minimum use of using a GPU with lightening.ai, not to mention prices at twice the cost of mystic.ai, or more. Pricing is higher for low usage on mystic.ai, but still quite affordable, and there is no low usage cost for a shared GPU.