feat: add voice cloning support to TTS endpoint

[rui8616]

Summary

• Add ref_audio (base64 data URI / URL) and ref_text parameters to /v1/audio/speech endpoint
• Pass voice cloning params to mlx-audio's native generate() ICL path instead of requiring a separate generate_voice_clone() method
• Auto-truncate reference audio to 15s max to prevent ICL context overflow that produces noise

Test plan

• 29 unit tests pass (including 7 voice-clone specific tests)
• Verified with real Qwen3-TTS-12Hz-1.7B-Base model — voice cloning produces usable audio
• Normal TTS (without ref_audio) unaffected

[ethannortharc]

Thanks for picking this up — the approach of using mlx-audio's native generate(ref_audio=, ref_text=) ICL path is the right call. A few things I noticed:

SSRF risk in _decode_ref_audio

urllib.request.urlretrieve will fetch any URL the user provides, including internal network addresses (http://169.254.169.254/, http://localhost:...). If the server is only used locally this is low risk, but worth considering if anyone exposes it on a network. Could restrict to validated external URLs or drop URL support and only accept base64.

Request body size limit

The audio file upload endpoints (/v1/audio/transcriptions, /v1/audio/process) have MAX_AUDIO_UPLOAD_BYTES protection via _read_upload(), but the JSON-based /v1/audio/speech endpoint has no equivalent body size limit. With ref_audio now carrying full audio files as base64, it might be worth adding a size check before base64.b64decode — something like rejecting payloads over a reasonable threshold. This isn't specific to your PR (the gap existed before), but since you're adding the first large-binary-in-JSON field, it would be a good place to address it.

15s truncation only works for WAV

_truncate_ref_audio catches wave.Error for non-WAV formats and skips truncation. If someone sends a long mp3/ogg as ref_audio, it would still overflow the ICL context. Could use mlx_audio.codec.load_audio (or similar) to handle more formats, or at least document the WAV-only limitation.

Minor

• ref_text or "" converts None to empty string — some models might treat these differently
• Temp file suffix is always .wav regardless of actual format — probably fine since load_audio detects from content, but worth noting

[NetLops]

Hi! Thanks for the great implementation. I'm really excited to try out this voice cloning feature. Any updates on when this PR might be merged?