Voices for narration
Esta página aún no está traducida; aquí tienes la versión en inglés.
Explainers, course videos and AI shorts are narrated by a voice. Reelfold picks one this computer can actually use, so a video never stops because a paid voice has no key.
Which voice is used
Section titled “Which voice is used”In order:
- Your persona’s voice, if you set one (
tts.engineinpersona.local.yaml, for example your own cloned voice). - OpenAI (
gpt-4o-mini-tts), only when an OpenAI key is set in Settings › AI. About $0.015 per minute of narration. - The free voices, which run offline on this Mac (Apple silicon). English uses Kokoro (
af_heart, witham_michaelandbf_emmaas alternatives). Chinese uses Qwen3-TTS once you have downloaded it (see below).
On autopilot, a project set to OpenAI with no key uses the free voice and says so in What the AI decided. With Ask me first it asks: Use the free voice or Add your OpenAI key. With no free voice either, it asks for the key.
The free voice download
Section titled “The free voice download”The free English voice (Kokoro) is a one-time download of about 330 MB. It starts the first time a video needs it, or from Settings › Advanced › Downloads. Until it is there, the voice step says “The free voice isn’t downloaded on this computer yet” with Download the free voice; press Try again when the download is done.
The free voices work offline and in the Mac App Store edition. Intel Macs and Windows don’t have them: narration there needs an OpenAI key (or another voice you set in the persona).
The free Chinese voice
Section titled “The free Chinese voice”Chinese versions are narrated by Qwen3-TTS (Apache-2.0), an optional one-time download of about 3 GB in Settings › Voice › Chinese voice. It reads tones and English terms naturally. You pick one of five voices there, and you can listen to each before downloading:
| Voice | Sound |
|---|---|
| Serena | Warm and calm (the default) |
| Vivian | Bright and young |
| Uncle Fu | Mature, low male voice |
| Dylan | Male, Beijing accent |
| Eric | Male, Sichuan accent |
Once it is downloaded, it narrates every Chinese version. Until then:
- Chinese versions use English narration with Chinese subtitles. When you ask for a Chinese version, the plan card says so and offers Download the Chinese voice; versions written after the download finishes are narrated in Chinese.
- The older built-in Mandarin voice (Kokoro,
zf_xiaobei) is off by default because it sounds robotic. To use it until Qwen3-TTS is downloaded, turn on Use the built-in Chinese voice until then in Settings › Voice. Once Qwen3-TTS is there, it is always used.
With an OpenAI key, narration uses OpenAI’s voices instead, in Chinese too.
English terms in Mandarin narration
Section titled “English terms in Mandarin narration”A Chinese course is full of English terms: MCP, API, Tool Calling, USB-C. Qwen3-TTS reads them by itself, with no dictionary. The older built-in Mandarin voice (Kokoro), if you turned it on, reads Chinese as Chinese and each run of English as English:
- Acronyms in capitals are read letter by letter: “MCP” is M-C-P.
- Words it knows are read as words: “Tool Calling”, “server”.
- CamelCase names are read in their pieces: “LangChain” is Lang Chain, “OpenAI” is Open A I.
- A word it doesn’t know at all is spelled out rather than skipped.
- Numbers on their own stay with the Chinese and are read in Chinese.
Tips for rare words and long sentences
Section titled “Tips for rare words and long sentences”- If a rare word is misread, reword the sentence or use a more common word. Writing the word the way it sounds also works, but the captions follow the script, so they will show that spelling too.
- Write acronyms in capitals when they should be spelled out (“API”, not “Api”).
- Spell numbers the way they should be spoken in English narration (“zero point five one”, “F P thirty-two”); captions turn them back into digits.
- Keep sentences short. A very long sentence is split to fit what the voice can read in one breath. If the voice still leaves something out, Reelfold says which scene and which sentence (see below).
When the voice stops
Section titled “When the voice stops”Every take is transcribed and compared with the script. A take that drops a sentence is made again, and a line that keeps failing is tried in two halves. If it still fails, the run stops with a plain message:
| Message | What to do |
|---|---|
| The voice kept leaving out a sentence in scene N: “…” | Reword or split that sentence in the script, then Try again |
| The voice could not read scene N, even in two halves: “…” | Same: simpler wording, shorter sentences, a rare word written as it sounds |
| There’s no OpenAI key in Reelfold, so this step couldn’t run | Use the free voice, or add the key in Settings › AI |
With the Claude Code skill
Section titled “With the Claude Code skill”tts: engine: kokoro # or qwen, openai, clone, edge, openai-compatible, elevenlabs zh_local: true # optional: Kokoro's Mandarin voice until Qwen3-TTS is downloadedtts_engine: auto (the default) follows the same order as the app; a Chinese line on the free voice goes to Qwen3-TTS when it is installed. Your own cloned voice (tts.clone, Qwen3-TTS on Apple silicon) also reads English terms inside Chinese. Engine details: Sound.