Higgsfield’s standalone audio is about one thing: voice. It turns text into spoken audio, lets you clone a voice from a sample, and lets you revoice an existing clip with a different speaker. That is the whole audio surface for a normal user, and it is worth saying plainly up front: there is no standalone music or sound-effects generator here. Higgsfield does have music and SFX models, but they exist only inside its game-generation pipeline and are not meant for general use. If you want a soundtrack, you either use a video model’s built-in native audio or bring your own.

This article covers the voice tools: the one call behind them, the two text-to-speech engines, how voices and cloning work, revoicing a video, and how voice fits into the rest of Higgsfield.

The one call behind spoken audio

Spoken audio comes from one tool, generate_audio:

  • a model (a text-to-speech engine),
  • a prompt — the words to speak,
  • a voice — a voice_type (preset or element) plus a voice_id,
  • optional tuning: speech rate, loudness, pitch, and output format.

It submits an async job and returns an audio file (WAV by default, or MP3 and others). One call produces one take, so a multi-line narration is several calls with the same voice.

Two text-to-speech engines

Higgsfield offers two general TTS models, and you pick by how much control you want over the engine:

  • seed_audio (ByteDance Seed Audio 1.0) — the default and the workhorse. It has the richest tuning: speech_rate (−50 to 100), loudness_rate (−50 to 100), pitch_rate (−12 to 12), plus output format (wav / mp3 / pcm / ogg_opus) and sample_rate up to 48 kHz. It can also take an audio reference to shape the delivery.
  • text2speech_v2 — a router to a named engine through its variant parameter: ElevenLabs, MiniMax, Seed Speech, Vibe Voice, or Cozy Voice. Reach for this when you specifically want one of those engines’ voices; otherwise seed_audio is the sensible default.

Both take the same voice_type + voice_id pair to choose who is speaking.

Voices: presets and your own

Voices come from list_voices, which returns two kinds:

  • Preset voices (voice_type: "preset") — a built-in library of ready voices.
  • Custom voices (voice_type: "element") — voices you created in your workspace, including clones.

Each voice has a voice_id, its voice_type, and a preview URL so you can hear it before committing. You pass the exact (voice_id, voice_type) pair to the audio models. The rule that keeps a narrator consistent across a whole video is simple: reuse the same pair on every line. That is exactly what the narrated-explainer workflow does — pick the voice once, then voice every block with it.

Cloning your own voice

To narrate in a specific voice — yours, a brand voice, a character — you clone it. In an Apps-UI client the create_voice tool opens a widget where you record or upload a clean speech sample of roughly 10 seconds to 3 minutes; the widget confirms the audio and creates the voice end to end. Cloning charges a one-time clone fee and runs asynchronously, so a fresh clone is usually still processing for a short while. Once it reports ready, it shows up in list_voices as a custom voice, and you use it with voice_type: "element" just like any preset.

If you already have a confirmed audio upload and need no widget, the backend create_voice_from_confirmed_audio does the same from an uploaded, confirmed audio id.

Revoicing a video

Separate from generating fresh narration, voice_change replaces the spoken voice in an existing clip with a different one, keeping the original timing and visuals and re-merging the new audio. You give it the source video and a target voice (voice_id + voice_type); the output dimensions come from the source. Use it to swap a placeholder narrator for a final one, or to try the same clip in several voices without re-rendering the video.

Tuning a take

Two habits produce natural narration:

  • Perform with parameters, not punctuation. Delivery comes from speech_rate, loudness_rate, and pitch_rate — not from stage directions in the text. Write the plain words you want spoken; do not add bracketed cues like “(excited)” or “(pause)”, which the model would try to read aloud.
  • Keep lines the right length. For narration synced to video, size each line to the clip it plays over, and nudge speech_rate up if a take runs long. Spell numbers out (“twenty percent”) so they are read the way you mean.

The call in practice

Through a Higgsfield-connected assistant, first hear the voices, then generate:

Show me the narrator voices, then read this line in the one I pick: “The blank edge is not ignorance — it is an invitation to go and look.”

The assistant calls list_voices (with previews), you choose one, and it voices the line. Conceptually:

generate_audio
  model: "seed_audio"
  voice_type: "preset"          # or "element" for a cloned voice
  voice_id: "<id from list_voices>"
  prompt: "The blank edge is not ignorance, it is an invitation to go and look."
  # optional: speech_rate, loudness_rate, pitch_rate

A short spoken line like this priced at about 0.4 credits when this guide was written, so voice is inexpensive; cloning adds its one-time fee.

Native audio in video is a different thing

Do not confuse the voice tools with the native audio many video models produce. Models like Veo, Kling, Seedance, and Gemini Omni can generate sound — including dialogue — baked into the clip as it renders, controlled by a generate_audio or sound toggle on the video model. That is separate from the generate_audio tool described here, which produces a standalone voice track you add in post. For a narrated explainer you use the standalone voice tool per block; for a self-contained talking clip you might let the video model’s native audio carry it.

How voice fits with the rest of Higgsfield

  • Narrated explainers — one voice reads each block; see Turn a Script into a YouTube Video, which voices every block with seed_audio.
  • Talking-head and UGC ads — Marketing Studio pairs an avatar with a voice.
  • Revoicing and dubbing — voice_change swaps a clip’s speaker; a separate dubbing tool handles translating a video’s speech into another language.
  • Your own brand voice — clone once, then reuse it everywhere as a custom (element) voice.

Tips

  • Reuse one voice across a whole piece for a consistent narrator; switching voice ids mid-video breaks the illusion.
  • Clone for a signature voice, then select it as an element voice on every take.
  • Pick text2speech_v2 with a variant when you want a specific engine (ElevenLabs and the others); stay on seed_audio otherwise.
  • Tune with the rate parameters, and keep the spoken text free of bracketed cues.
  • Do not look for a music button. For a soundtrack, use a video model’s native audio or add your own track; standalone music and SFX are not part of the general toolset.

Recap

Higgsfield audio is voice: generate_audio turns text into a spoken take through seed_audio (rich tuning) or text2speech_v2 (ElevenLabs, MiniMax, and other named engines), using preset or cloned voices from list_voices. You can clone your own voice from a short sample and revoice an existing clip with voice_change. It is cheap per line, performance comes from the rate parameters rather than text cues, and there is no standalone music or SFX — that is the one gap to plan around.