Neural Narration Guide: Give Every Book a Voice

StoryCodex ships with three neural text to speech engines, and they are one of the reasons people stay: Piper, Kokoro, and Supertonic. Each is a different voice family with its own strengths, and you can switch engines between books, between chapters, or mid-sentence whenever the mood changes. All three run entirely on your device after a one-time voice download: no streaming, no account, no subscription. If you want the technical breakdown of models and licenses, the engine comparison covers it; this guide is about how the voices work, which to pick, and getting listening.

How a neural voice actually works on a phone

Every text-to-speech system solves the same two problems. First it has to decide what to say: which speech sounds, in which order, with what pitch, stress, and timing. Then it has to render that plan into an actual waveform your speaker can play. Older engines handled the first part with pronunciation dictionaries and hand-written rules, and the second part by gluing together short recordings of a human voice. Neural engines learn both halves from data, which is why the output carries cadence and natural breath instead of sounding assembled.

The three engines divide that work differently. Piper is a VITS model, an end-to-end network that maps phonemes directly to waveform; a compact rule-based component called espeak-ng converts written English into phonemes before the model sees them. Kokoro follows a similar phoneme-first shape but adds a style-conditioned model in the StyleTTS 2 family, plus a bundled US English lexicon for words espeak-ng would guess at. Supertonic goes furthest: its pack contains four separate int8 models, a text encoder, a duration predictor, a vector estimator, and a vocoder, and each utterance passes through an eight-step generation process before audio comes out.

All three packs ship in int8 form, and that single detail is what makes offline neural audio practical. Quantization stores each model weight as an 8-bit integer instead of a 32-bit floating point number, which shrinks the model roughly fourfold and lets it run at better than real-time speed on ordinary phone processors. The cost is a small amount of numeric precision, and in practice what you hear is the character of the model rather than the compression. This is also why the downloads are so modest: 79 to 141 MB buys you an entire voice engine, roughly the size of a two-minute video.

Underneath, all three engines run through sherpa-onnx on ONNX Runtime. StoryCodex picks an execution backend automatically: it prefers NNAPI, which lets Android route work to a GPU, NPU, or DSP where the device and engine support it, then an optimized CPU path called XNNPACK, then plain CPU as the universal fallback. If a backend fails or crashes during model load, the app quietly steps down to a safer one rather than breaking narration. None of this needs configuring; it is why the same pack runs on a budget phone and a flagship without drama.

Choosing by what you actually read

Fiction puts unusual demands on a speech engine, and StoryCodex does some quiet work on your behalf. Before synthesis, the text passes through normalizers that turn things like Roman numerals and acronyms into speakable words, so 'Chapter IV' is read as a number rather than spelled out letter by letter. Invented names are still the hard part: the engine phonemizes them as written English, which produces a consistent approximation rather than a native pronunciation. Consistency matters more than perfection over a 1,500-chapter serial, because your ear learns the voice's version of the name.

The listening experience

Open the player and the full toolkit is there: narration speed presets from 0.75x to 2.0x, a sleep timer (off, end of chapter, or 5, 15, 30, or 60 minutes), auto-next-chapter, and voice switching on the fly. A compact player bar follows you through the app, and the full-screen Now Reading view works with lock screen controls, Bluetooth headset buttons, and Android Auto. The hands-free listening guide walks through commuting and workout setups.

The detail readers love most is sentence highlighting. As the narration speaks, the current sentence lights up in the text, so you can follow along and drop back in at any moment. Pause mid-chapter, tap a word, and the voice continues from exactly where you tapped. It is the same engine that lets you turn any web novel or EPUB into an audiobook, except the text stays on screen if you want it.

There is also a fourth option that needs no download at all: your phone's built-in system voice. It sounds more mechanical than the neural engines, but it works instantly, which makes it handy for a quick preview or for short pasted text while a voice pack downloads. Most readers try it once, hear the difference a neural model makes on the same paragraph, and never go back. If you are coming from a stock OS voice and want the full upgrade path, see replace robotic TTS with natural AI voices.

Storage, battery, and other practicalities

Voice packs live in app storage as a handful of files: the model itself, a token table or lexicon, and in Piper and Kokoro's case the espeak-ng data that handles phonemization. Removing a pack frees its full size and never touches your books, progress, or Story Codex data. You can also keep all three installed at once; the whole set costs under 350 MB, less than a single downloaded movie, and switching between them is instant.

Battery use is gentler than you might expect because synthesis is fast and the app works ahead of you. While a sentence plays, a background prefetcher synthesizes upcoming pieces and caches the audio to disk, so pausing and resuming replays finished audio instead of regenerating it. For long trips you can pre-render whole chapters ahead of time, which shifts the compute cost to a convenient moment; after that, playback is just ordinary audio decoding. The sleep timer and auto-next-chapter round out a setup that is comfortable leaving on a nightstand.

StoryCodex is free to try on Google Play and the App Store. Import a serial, pick a neural voice, and your library works fully offline, no account required.