Feature idea: built-in real-time "voice-to-voice" conversion using Piper TTS voices

[Tamila-2017]

Feature idea

It would be great if Handy could offer an optional real-time "voice-to-voice" mode: take the microphone input, run it through Handy's own STT, then immediately synthesize the same text with a chosen Piper voice (see the voice list/samples at https://rhasspy.github.io/piper-samples/) and output it to a virtual microphone device - effectively an offline voice changer. Useful for anonymized calls (Telegram, Discord, etc.), accessibility, or just fun/privacy use cases.

What we already tried

We built exactly this as an external pipeline around Handy, using the "External Script" paste method: Handy's STT result gets passed to a script that runs it through Piper TTS (CPU, fully offline) and plays the result into a PulseAudio virtual sink (module-null-sink) that another app (Telegram) picks up as its "microphone".

It technically works, but we hit two serious problems that would probably need first-class support inside Handy (rather than shell-script glue around it) to fix properly:

• Instability: audio sometimes played back correctly within about 2 seconds, but other times a completely different, much older phrase (spoken many minutes earlier) would come out instead - as if a backlog/queue built up somewhere in the virtual audio device and never drained.
• Latency: even when "working", real end-to-end delay was often 15-30+ seconds from finishing speaking to the synthesized voice reaching the other side of a call - far more than the roughly 2 seconds our own script-level timing measurements accounted for. So the bottleneck was in the OS audio plumbing (PulseAudio null-sink plus monitor), not in Handy's own STT or in Piper's synthesis - both of those were consistently fast, well under 1 second each, according to our logs.
  We eventually gave up on the external-script approach specifically because of this delay/stability problem. If this became a built-in Handy feature with a proper native audio output path (instead of routing through a null-sink and its monitor), it would likely avoid this whole class of bug.

To be clear, this is a request for generic offline TTS voices (like Piper's existing voice list), not for cloning any specific real person's voice.

[cjpais]

To be honest, this is not Handy's purpose.

[Doroshenko-T]

Of course I agree with you. But this could be a great addition to it :-)