Synthesis Pathway: When Do We Need the Encoders?

codec_encoder.fp16.onnx and speaker_encoder.fp16.onnx are only required when the engine has to extract speaker / codec features from a raw WAV in real time. Every other path is "cold inference" — it consumes text, instructions, a built-in speaker ID, or a pre-computed JSON.

Cold-inference pathways (no encoder needed)

Only talker (GGUF), predictor (GGUF), and decoder (ONNX) are touched.

Model type	Pathway	Notes
Base (0.6B / 1.7B)	Standard TTS	Text only; speaker_embedding / ref_codes are null.
Base (0.6B / 1.7B)	Clone from JSON	--reference foo.json — features already cached.
Custom (0.6B / 1.7B)	Standard TTS	Default voice.
Custom (0.6B / 1.7B)	Custom voice	--speaker Vivian — embeddings loaded from embeddings/.
Design (1.7B)	Standard TTS	Text only.
Design (1.7B)	Voice design	--instruct only.

The one path that still needs encoders

Model type	Pathway	Why
Base (0.6B / 1.7B)	Clone from WAV	Real-time extraction of acoustic codes + speaker embedding from --reference foo.wav.

custom and design models do not support clone at all (the loader rejects --reference).

Cleanup tip

If you only ever use cold-inference pathways, both speaker_encoder.fp16.onnx and codec_encoder.fp16.onnx can be deleted from any models/<name>/ directory. Saves ~130–140 MB per model. The runtime tolerates their absence.