litellm - 💡(How to fix) Fix [Bug]: router.aspeech() bypasses async_function_with_fallbacks — TTS requests have no retry or failover

What happened

Router.aspeech() bypasses async_function_with_fallbacks and calls litellm.aspeech directly. As a result:

• No retry on failure — if the selected deployment returns an error, the request fails immediately with no retry attempt.
• No failover to other deployments — if you have multiple TTS deployments under the same model_name, only the first deployment selected by async_get_available_deployment is ever tried. Any error surfaces to the caller.

This is inconsistent with every other router method (acompletion, aembedding, arerank, atext_completion, etc.) which all route through async_function_with_fallbacks.

Root cause

litellm/router.py — aspeech() (around line 3542):

async def aspeech(self, model, input, voice, **kwargs):
    ...
    deployment = await self.async_get_available_deployment(...)
    self._update_kwargs_before_fallbacks(model=model, kwargs=kwargs)
    ...
    response = await litellm.aspeech(...)   # ← direct call, no fallback loop
    return response

Compare to aembedding or arerank, which set kwargs["original_function"] and call await self.async_function_with_fallbacks(**kwargs). aspeech has no such wrapper.

Impact

Any operator running multiple TTS providers (e.g. OpenAI TTS + Azure TTS) for redundancy gets no automatic failover. A single deployment outage causes 100% of TTS requests to fail even if a healthy backup is configured.

Expected behavior

aspeech() should route through the same retry/fallover mechanism as other router methods — either by:

1. Adding an _aspeech inner method that implements the actual litellm.aspeech call, setting kwargs["original_function"] = self._aspeech, and then calling await self.async_function_with_fallbacks(**kwargs), or
2. Wrapping the direct call in retry logic consistent with the rest of the router.

Related

This is a separate issue from #27390, which covers the missing _update_kwargs_with_deployment call (TTS spend tracking bug). Both need fixing independently.

LiteLLM version

Reproduced against main branch.