Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

5.3 KB
Raw Blame

Speech: turn a script into a voice

Lip-sync it adds scene ambience and sound effects by default on LTX-2.5. Describe the setting in the scene prompt, or use Scene sound description for specific sounds. Turn Add scene sound off for voice only. This adds one sound-generation pass with the installed LTX model. The clean voice still drives the lips; the bed is mixed afterward and lowers during speech. Effects follow the scene description; exact action timing is not guaranteed. Add optional music on the Editor timeline.

Open Audio → Speech to turn a script into a spoken WAV file on your Mac. Speech sits alongside Compose and Music video. Choose a preset voice, a saved character's voice or a clip you add. The finished file appears in Outputs and can go straight to Lip-sync.

Install Speech {#install}

Speech needs a one-time download of about 6 GB for the speech engine and two voice models.

  • What it needs: an Apple Silicon Mac with 16 GB of unified memory or more, and about 8 GB of free disk for the install.
  • How: open Audio → Speech. The form stays visible but greyed out until installation finishes. Use the in-app install card or click Install Speech (~6 GB), which appears in place of Generate.
  • Pause and resume: click Stop to pause the download. Start the install again to resume it.
  • Licence: the models use Apache-2.0. There is no licence screen.

After installation, Speech works offline.

Write your script {#script}

Enter up to 2,000 characters. Each line is spoken in turn, with a short pause between lines. Punctuation shapes the delivery: try !, ? or … where you want emphasis, a question or a pause.

Below the box, “About N s of speech” estimates the length. The app also shows whether your script fits one lip-sync clip: about 10 seconds, or roughly 25 words. Split a longer script into shorter scripts for individual clips.

Choose a voice {#voice}

  • Preset voice: choose from nine model voices. Ryan and Aiden are native English speakers. Serena, Vivian, Uncle Fu, Dylan and Eric are native Chinese speakers; Ono Anna is Japanese and Sohee is Korean. All can speak English, with an accent for the non-native English voices. These are the model's own voices, not real people.
  • A character: use the voice sample a saved character was trained with in the Train tab. Only characters with a voice sample appear in the list. Type the words spoken in that sample for a closer copy of the voice; Phosphene remembers them for that character.
  • A clip you add: add 5–20 seconds of one person speaking clearly, without music. Use WAV, MP3, M4A or FLAC. You can also type the words spoken in the clip for a closer copy of the voice.

Permission to use a voice {#consent}

Each time you use A character or A clip you add, tick This is my voice, or I have the speaker's permission to use it. The confirmation is saved with the file.

Only clone a voice with the speaker's permission. Never use a cloned voice to deceive. Phosphene's acceptable-use policy forbids impersonation.

Adjust the delivery {#controls}

For preset voices, choose a Delivery: Natural, Warm, Excited, Deadpan, Dramatic, Whisper, Sad or Angry. Or type your own direction below the styles, such as “tired, like it's 3 a.m. and the render just failed”. Typed direction takes priority over the selected style.

Character and clip voices have no delivery styles. Their delivery follows the clip's mood and your script's punctuation.

Control What it does
Speed Set the pace from 0.70× to 1.40×. The pitch stays the same
Advanced → Language Choose English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish or Italian
Advanced → Seed The same seed, text and voice give the same take. Leave -1 for a random seed

Generate and use your speech {#generate}

Click Generate. Speech joins the same queue as videos and songs, so it waits if a video is rendering. Once it starts, a line takes a few seconds. The finished speech plays right below Generate.

  • Click Lip-sync it to open Lip-sync with the audio loaded. Add a picture of the face, then click Generate.
  • Click Use this audio for… to choose Lip-sync, Lip-sync with Hailuo H3 when installed, Music video, Soundtrack for a video or Cover it.

Find the file in Outputs using the Audio filter. It is a 24 kHz mono WAV. A small .json file beside it records the text, voice and seed.

For agents and scripts, use POST /speech/generate; see the API doc.

When something goes wrong {#troubleshooting}

What you see What to do
Installation failed Click Install Speech (~6 GB) again. The download resumes and keeps finished files
“Stopped before the Mac ran out of memory” Close other apps and browser tabs, then try again
A cloned voice sounds off Use a cleaner 10–15 second clip with no music or background noise, and type its words
A preset sounds too flat Choose a Delivery or type your own direction

Credits {#credits}

Speech uses Qwen3-TTS by the Qwen team (Apache-2.0), running on Apple's MLX through mlx-audio (MIT).