markdown
Offline Windows TTS narration studio (Kokoro). Not affiliated with narravoice.com.
A local, offline text-to-speech narration studio for turning story files into audiobooks — no cloud APIs, no per-character billing, no subscription.
Two engines in one app Kokoro — fast, fully local, lightweight. Qwen3-TTS — higher expressiveness (optional through Python path) Use either (or both) in the same project: same editor, presets, chunking, and export pipeline.
NarraVoice runs entirely on your machine. The default engine, Kokoro, is CPU-friendly out of the box, so it works well even on older or integrated GPUs. An optional second engine, Qwen3-TTS, is available for higher-quality synthesis when you have the hardware (or patience) for it.
- Fully offline — both TTS engines run locally; nothing is sent to a cloud API.
- CPU-first — optimized to run well without a dedicated GPU.
- Dual-engine — mix Kokoro and Qwen3-TTS voices in the same project, even the same story.
- Per-line voice control — a preset/gutter system lets you arm a different voice, pitch, rate, or volume for any stretch of text, and it holds until the next marker. Great for multi-character dialogue or shifting a single character's tone mid-scene.
- PDF/DOCX/TXT ingestion — drop in a story file and NarraVoice chunks it automatically at sentence boundaries (~2500 characters per chunk) for editing and rendering.
- Full pipeline — edit → preview → render each chunk → merge into one finished MP3 audiobook.
- Combine external audio — Doesn't just combine audio from NarraVoice, but if you have external audio files you'd like to combine, NaraVoice can handle that too.
- Built-in visualizer — waveform, pitch (Hz), and energy (RMS) view for any rendered audio. Useful for spotting flat delivery or odd chunk-boundary artifacts. View Audio files not generated by NarraVoice also.
- Pronunciation control — Smart IPA overrides for Kokoro (with homograph handling via a built-in dictionary + eSpeak NG fallback), plus a global substitution list for both engines.
| Feature | NarraVoice (Local & Offline) | Cloud-Based TTS Engines |
|---|---|---|
| Cost | 💸 100% Free — No subscriptions, no credit cards. | 💳 Per-character billing — Gets expensive fast for long books. |
| Privacy | 🔒 Fully Private — Text and audio never leave your PC. | 🌐 Cloud Logging — Sent to third-party corporate servers. |
| Hardware | 💻 CPU-Optimized — Designed to run smoothly on standard laptops. | ☁️ Server Dependent — Requires internet; subject to server outages. |
| Dialogue Control | 🎭 Per-Line Presets — Gutter colors track multi-character voice changes. | 🧩 Manual Splitting — Hard to manage shifting character tones in one file. |
| Fine-Tuning | 🛠️ Smart IPA & Local Dictionaries — Total manual control over heteronyms. | 🤖 AI Guesswork — Often mispronounces names/heteronyms with no bypass. |
A typical install looks like this (adjust paths to your own setup):
YourInstallFolder\NarraVoice config.json narration_config.json substitutions.json voice_preferences.json projects models\kokoro\Kokoro-v1.0.onnx models\kokoro\voices\ ← *.bin voice files models\Qwen\qwen_server.py ← optional models\Qwen\Qwen3-TTS-1.7B-real\ ← optional model tree espeak\ ← optional full eSpeak NG copy
{
"base_dir": "C:\\Path\\To\\NarraVoice",
"app_dir": "C:\\Path\\To\\NarraVoice",
"models_dir": "C:\\Path\\To\\NarraVoice\\models",
"voices_dir": "C:\\Path\\To\\NarraVoice\\models\\kokoro\\voices",
"projects_dir": "C:\\Path\\To\\NarraVoice\\projects",
"kokoro_model_path": "C:\\Path\\To\\NarraVoice\\models\\kokoro\\Kokoro-v1.0.onnx",
"qwen_server_script": "C:\\Path\\To\\NarraVoice\\models\\Qwen\\qwen_server.py",
"espeak_path": "C:\\Program Files\\eSpeak NG\\espeak-ng.exe",
"device": "cpu"
}Place config.json next to the NarraVoice executable and update the paths to match your install. eSpeak can stay in Program Files; only copy the full eSpeak NG folder under espeak\ if you want a portable install.
On first use: Tools → Voice Manager / Download Voices. Missing English voice .bin files download automatically. Defaults show en-us and en-gb only — use Show All for other languages.
Qwen3 unlocks a second, higher-quality voice engine at the cost of extra setup and slower CPU-only synthesis.
A note on "optional": Qwen3-TTS as a whole is optional — if you don't want it, skip this section entirely and NarraVoice works fine on Kokoro alone. But if you do choose to use Qwen3, every step below stops being optional and becomes required. There's no partial setup.
- Install Python. Development and testing were done on 3.13.14 — other versions haven't been tested and may or may not work.
- Download the Qwen3-TTS model and server script from [source/link] and place them under
models\Qwen\(or pointqwen_server_scriptat wherever you put them). - The first Qwen preview starts a local server at
http://127.0.0.1:8765automatically.
New Project → pick a story file (.txt, .docx, .pdf) → edit a chunk → Preview. Once all chunks in a story are rendered, Merge produces one MP3 under audiobooks\.
💡 Import Tip (Subheaders & Line Breaks): Because NarraVoice splits chunks strictly at sentence boundaries, imported header lines or lines without punctuation (e.g., ., ?, !) can sometimes merge with the line below them. If this happens, simply place your cursor at the break in the editor and press Enter to split them. NarraVoice does not auto-insert periods, ensuring your manual line splits match your gutter presets and intonation perfectly.
| Task | How |
|---|---|
| Add pronunciation | Right-click word → Smart IPA (Kokoro only) |
| Insert a pause | Silence button, <sil:500ms>, or a blank line |
| Hear current chunk | Preview |
| Hear selected text | Highlight text → Preview |
| Replay last preview | Play |
| Fix flat question intonation | Use bf_alice or af_jessica, or blend them |
| Manage which voices show | Tools → Voice Manager / Download Voices |
| Set a Qwen style note | Select a Qwen voice → Instruct… button appears below the voice dropdown |
| Fix odd/silent audio | See Phantom Sounds, below |
Save a Voice + Rate + Pitch + Volume combination as a named preset (each gets a gutter color). Arm a preset in the left margin and it applies from that line until the next marker — this is how you get multiple voices, or the same voice in different moods, within one story.
For Qwen voices only, a preset can also carry an Instruct — a short style note (e.g. "speak angrily") that's saved with the project and applied every time that preset renders. There's a second, separate session Instruct for quick one-off experiments that is not saved between app restarts.
Render priority for a given line:
- Preset Instruct (if the line is marked with a preset that has one)
- Session Instruct (if set)
- No instruct
Kokoro voices always ignore Instruct — pitch/rate/volume are the only controls there.
- Kokoro: right-click any word → Smart IPA for a homograph-aware suggestion (backed by a built-in dictionary, falling back to eSpeak NG), or add it to the global substitution list.
- Qwen3: doesn't use IPA. Instead, spell the word the way it should sound (e.g. write "reed" if you mean the present tense of "read").
- Don't combine a global substitution and an inline IPA override on the same word — pick one.
- Heteronyms (read/read, lead/lead, live/live, etc.) aren't resolved automatically — text-to-speech in general still struggles with these. NarraVoice gives you manual override tools rather than guessing; expect the occasional line to need a one-off fix.
Phantom sounds — usually an empty segment caused by stray quote characters, runs of punctuation, or too many blank lines. Prefer the Silence button or a single <sil:...> tag over blank lines for pacing.
Pitch — stay roughly within −6 to +6 semitones; above about +5 tends to sound harsh. Pitch changes overall pitch only — it's not a substitute for a voice with better natural prosody.
- Most Kokoro voices have limited question-intonation (prosody) —
bf_aliceandaf_jessicaare the strongest, or blend them. - Cancel during synthesis isn't instant, and occasionally doesn't stop an in-flight Kokoro segment at all.
- Qwen3 requires a separate Python + server + model install, and is noticeably slower than Kokoro on CPU-only hardware.
- Multi-voice dialogue within a single sentence often sounds unnatural — presets work best when switching voices at line/paragraph boundaries, or for shifting one narrator's emotional tone.
- Session Instruct is intentionally not persisted across restarts.
- Heteronyms need manual pronunciation overrides — see Pronunciation, above.
See the User Guide for the complete editor reference, keyboard shortcuts, and every setting in depth. Also see the context document for more technical information.
NarraVoice is licensed under the MIT License — see LICENSE.
Third-party packages and user-installed models are covered in THIRD_PARTY_NOTICES.md. Kokoro and Qwen3-TTS model weights are Apache 2.0 and are not redistributed with this project; download them separately.