Implements a sliding-window repetition history with FIFO eviction (`_ChannelWindow`) used to penalize repeated token sequences during autoregressive generation. The class is self-contained, isolated from the rest of the engine, and can be reused in any autoregressive text or audio generation system that requires repetition penalty. It exposes a clean interface for resetting, pushing, and checking the history, and supports both single-path and batched generation.

Provides a lazy-loaded ONNX inference wrapper for the speaker encoder (x-vector model). It handles model loading, input preprocessing, and embedding extraction from a reference audio clip. The module is fully decoupled from the rest of the TTS pipeline and can be extracted as a standalone speaker embedding service for any voice cloning or speaker verification task. It uses ONNX Runtime exclusively, with no PyTorch dependency.

Contains the babble guard heuristic (`babble_suspect`, `babble_prefer`) that detects early EOS or run-on artifacts in generated audio chunks and triggers re-generation with adjusted parameters. The logic is self-contained, operates on token count and log-probability statistics, and can be adapted to any autoregressive audio generation system that needs to detect and recover from truncated or looping outputs. Also includes reusable audio utilities like `join_audio_chunks`, `trim_and_fade`, and `pause_pad_samples`.

Implements a clean factory function `Vieneu(mode=...)` that returns the appropriate TTS engine class based on a string identifier (e.g., 'turbo', 'nano', 'standard', 'fast', 'remote'). It abstracts away the complex selection logic and allows users to switch backends with a single parameter. The pattern is well-structured and reusable for any multi-backend system where different implementations share a common interface (defined in `base.py`).