The Generate Voice panel makes everything you hear: speech from text, voice changes, sound effects, songs and background music, a soundtrack or foley for a video, dubbing into another language, and transcripts. It is the same generation panel as images and video, with its own set of model categories.
Set up
Most audio models run on a WaveSpeed key (Get API Key). A Pika key adds Pika’s own audio line plus hosted ElevenLabs, Google Lyria, Kling, MiniMax and Sonilo models — see Pika and HiggsField. Keys go in Settings — see Configuring API Keys.
Open the panel
Click Generate Voice in the bottom toolbar. The panel docks at the bottom of the board, like the image and video panels, and keeps your text, inputs and model when you close it.
Click the model pill in the bottom bar to choose a model; the category chips at the top of the list narrow it to one job. For speech models the prompt box is the text to speak (“Text to turn into speech…”). Models that work from a source file, such as a voice changer, show an optional prompt and take their input in the slot row.
Audio categories
| Category | What it does | Models |
|---|---|---|
| Text to Speech | Speaks your text | WaveSpeed: Chatterbox TTS, Zonos 2 Voice Clone, MiniMax Speech 2.8 HD, Gemini 3.8 Flash TTS, Gemini 3.8 Flash-Lite TTS. Pika: Pika Speech, ElevenLabs Multilingual V2, Turbo V2.5 and Text to Dialogue, MiniMax Speech 2.8 HD and Turbo. |
| Text to Audio | Sound effects, or voice and sound from one prompt | WaveSpeed: Seed Audio 1.0. Pika: Pika SFX, ElevenLabs sound effects, Lyria 2 Sound Effects, Lyria 3.5, Kling Audio, Sonilo sound effects, Seed Audio 1.0. |
| Speech to Speech | Re-voices a recording in another voice | WaveSpeed: Chatterbox Speech-to-Speech. Pika: ElevenLabs Multilingual STS. |
| Text to Music | Songs with vocals, or instrumental background music | WaveSpeed: Mureka V9 Generate Song and Generate BGM, Sonilo Text-to-Music, MiniMax Music 02, HeartMula Generate Music. Pika: Pika Music, Lyria 3 and Lyria 3 Pro, MiniMax Music 3.0, Sonilo Text to Music. |
| Music to Music | Re-styles an existing song | MiniMax Music Cover |
| Music to Text | Describes a track’s genre, mood and instrumentation | Mureka Describe Song |
| Video to Audio | Foley, ambience, sound effects or a soundtrack for a video | WaveSpeed: Hunyuan Video Foley, MMAudio V2, Kling Video-to-Audio, Sonilo Video Soundtrack. Pika: Pika Soundtrack, Kling Audio, Sonilo sound effects and scored video. |
| Video to Music | Scores a video | Pika: Sonilo Video to Music |
| Voice Isolation | Pulls the voice out of a noisy recording | Pika: ElevenLabs Voice Isolation |
| Voice Dubbing | Re-voices the speech in a clip in another language | Pika: ElevenLabs Voice Dubbing |
| Transcribe (Audio) / Transcribe (Video) | Turns speech into text | WaveSpeed: Whisper Transcribe and Whisper Transcribe Video. Pika: Whisper and ElevenLabs Scribe. |
A category chip only appears when one of your keys unlocks a model in it. Pika’s list comes from Pika’s own catalogue and changes as Pika adds models.
Text to speech
Type the text in the prompt box, pick a voice and click Generate.
| Model | Voice | Controls | Price |
|---|---|---|---|
| Gemini 3.8 Flash TTS | 30 preset voices, Kore by default | Style Instructions — tone, pacing, accent or emotion (“warm and unhurried, slight British accent”). Inline cues such as <laugh> or <sigh> are performed. | $0.05 per 1,000 characters |
| Gemini 3.8 Flash-Lite TTS | Same 30 voices | Same Style Instructions | $0.04 per 1,000 characters |
| MiniMax Speech 2.8 HD | A system voice or one you cloned on WaveSpeed, by Voice ID | Emotion, Speed (0.5–2), Volume, Pitch (−12 to 12), Language Boost, Pronunciation Dictionary, output format. Interjections such as (laughs) and (sighs) are spoken. | $0.10 per 1,000 characters |
| Chatterbox TTS | Optional Reference Voice to match | Exaggeration, Temperature, CFG / Pace | $0.03 per run |
| Zonos 2 Voice Clone | Clones the Reference Voice you attach (required) | Clean Background | $0.01 per run |
| Pika Speech | One of 76 voices in Voice Preset, or a Reference Audio to clone | Voice Consent Attested for cloning | $0.01 per minute |
The Gemini models bill the text plus the style instructions, with a minimum of one 1,000-character block — a short line costs the same as a 1,000-character one, and the estimate shows that minimum. Until you type anything, character-billed models show their rate as /1k chars.
Music
- Songs with vocals — Mureka V9 Generate Song, MiniMax Music 02 and HeartMula take the words in a Lyrics field under Advanced Settings. The prompt box sets the style (“dreamy synth-pop, female vocals”); HeartMula takes its style as Style Tags instead.
- Instrumental — Mureka V9 Generate BGM and Sonilo Text-to-Music work from a style prompt. Sonilo takes a length in seconds on the bottom bar (default 30).
- Pika Music — Mode chooses text to music, voice-conditioned text to music, or a music cover of a Source Audio. Duration is a dropdown on the bottom bar, from 10 to 360 seconds (default 30).
- MiniMax Music Cover (Music to Music) re-styles a Source Song from a prompt.
- Mureka Describe Song (Music to Text) reads a track and drops the description on the board as a note.
Sound for a video
Pick a Video to Audio or Video to Music model and put the clip in the video slot.
- Hunyuan Video Foley — synced foley and ambience, with an optional prompt.
- MMAudio V2 — sound effects and ambience from a prompt (required) and an optional negative prompt.
- Kling Video-to-Audio — a Sound Effects Prompt, a BGM Prompt and ASMR Mode.
- Sonilo Video Soundtrack — music and synced effects, with Sound Direction, an optional Music Prompt, Preserve Speech and Duck Under Speech.
- Pika Soundtrack — replaces the audio of a clip with a synchronised soundtrack; the picture and length are untouched.
Most of these return an audio card. Sonilo Video Soundtrack, Pika Soundtrack and Sonilo’s scored video return the clip itself with its new soundtrack, as a video card.
To score a shot inside a timeline, with a preview over the picture and takes to compare, use AI Sound Design in the Video Editor.
Dubbing, voice changing and isolation
- Voice Dubbing (ElevenLabs, via Pika) — put the source clip in the slot and type the target language in Target Lang under Advanced Settings.
- Speech to Speech — Chatterbox Speech-to-Speech converts the Source Audio into the voice of an optional Reference Voice.
- Voice Isolation (ElevenLabs, via Pika) — returns just the voice from a recording.
Audio inputs
Audio slots (a reference voice, a source song, a recording to convert) offer Audio on canvas, Select from media library and Upload audio, plus Record voice: click Start recording, then Stop, listen back, and click Use recording (or Re-record). The recording also lands on the board as an audio card, so you can reuse it.
When the model has several inputs you can name them in the prompt with @audio1, @image1 and so on — see Mentioning Canvas Media.
Where results land
As soon as you click Generate, a placeholder appears in the visible part of the board, above the panel — a short strip for audio. When the result comes back:
- speech, sound effects and music become audio cards;
- video-returning models (see above) become video cards;
- transcripts and song descriptions become notes holding the text.
The files are saved in the board’s media folder, and every run appears in Recent Generations — see Generation History.
In the Video Editor
- AI Sound Design runs the Video to Audio and Video to Music models on a selected video layer and places the result on an audio channel.
- Voice Channels and Transcribe Editor re-voices a take in its own voice with the voice-cloning models. Pika’s voices pick from preset voices rather than copying a recording, so they are offered in the Generate Voice panel but not there.
Where to go next
- AI Sound Design — foley and scores for a shot on the timeline.
- Voice Channels and Transcribe Editor — narration you can edit word by word.
- Audio Transcription — turn audio on the board into text.
- Pika and HiggsField — the provider behind most of the extra audio models.
- Video Generation — Omni models that take audio references.
Leave a Reply