The best AI models for voiceover
Narration and spoken lines. These are Call Sheet's picks for this job, in order, with the reason for each. Call Sheet is a free tool by Melissa Beery that ranks 101 image, video and sound models for a creative job. Its data was last updated on 20 October 2026.
1
ElevenLabs
ElevenLabs
The consent‑based voice library is the industry default for narration.
Eleven Music plus the voice tools.
- Cost
- Own platform, free tier, paid from ~US$5
- Where
- ElevenLabs's own site · check in app: Higgsfield, OpenArt, Runway
- Rights
- Full Enterprise Music carries IP indemnification for audio.
- Watch out
- The commercial license on self‑serve tiers excludes exactly the film, TV and radio work a producer would use it for.
How to ask itVoice: pick the voice, set Stability, then write the line with [audio tags] and punctuation. Music: genre + BPM + instruments + mood + era, then the arrangement in order, then lyrics or 'instrumental only'.v4 and v3 have no SSML break tag. Use ellipses and tags for pauses. Break tags work on v2 up to 3 seconds.
Write the arrangement in order with small words: start with, just, then, bring in. Mark the silences.
Open ElevenLabs in Call Sheet →
2
Hume Octave 2
Hume AI
Reads the meaning of the line and performs it. Acting directions in plain English.
Text to speech that reads the meaning of the line and performs it.
- Cost
- Own platform, commercial from Creator US$14
- Where
- Hume AI's own site
- Rights
- Starter (US$3) is non‑commercial.
- Watch out
- Eleven languages, voice design in English only, and Hume's own release notes admit speaker drift on long reads.
How to ask it[voice from the library, or none so the description designs one] + utterance text + description 'frightened, rushed' under 100 characters + speed 0.5 to 2.0Keep the description under 100 characters.
Use precise emotions, and add pacing words like 'measured' or 'deliberate pause'.
Open Hume Octave 2 in Call Sheet →
3
OpenAI text to speech
OpenAI
Cheap, steerable, and the one voice here with a written indemnity for business use.
API only.
- Cost
- API, about US$0.015 a minute
- Where
- OpenAI's own site
- Rights
- The OpenAI Services Agreement (13.1) defends business and API customers against third‑party IP claims over the Services.
- Watch out
- No web app, no cloning, English first, and a July 2026 bug thread on truncated final sentences.
How to ask itmodel gpt-4o-mini-tts + voice (marin or cedar) + input text + instructions: 'Speak in a warm, unhurried tone with a light Irish accent'Put direction in instructions, not in the text. The text is read as written.
Ask for wav or pcm when you need the file fast. mp3 is the default.
Open OpenAI text to speech in Call Sheet →
4
Gemini TTS
Google
Two speakers a pass in 70 plus languages, for very little. Still in Preview.
Single or two‑speaker reads in 70 plus languages, steered by a style prompt.
- Cost
- API, about US$0.015 a minute. Free tier trains on you
- Where
- Google's own site
- Rights
- Google's indemnity covers Gemini on Vertex AI, not the Gemini Developer API.
- Watch out
- Voice and tone drift between runs of the same text, per Google's own forum.
How to ask itStyle line, then the transcript: 'Say in a spooky whisper: "..."'. Two speakers: name each, give each a note, then Speaker1: and Speaker2: lines.Name the voice, then match the written tone to it. A deep voice asked to sound like a child mismatches.
Two speakers per pass at most, each named in the transcript.
Open Gemini TTS in Call Sheet →
5
Murf AI
Murf
A timeline editor for corporate narration, with full commercial rights from the first paid tier.
A studio‑style voiceover editor for explainers, e‑learning and ads. 200 plus stock voices on a timeline with music.
- Cost
- Own platform, Creator from US$19 a month billed yearly
- Where
- Murf's own site
- Rights
- Full commercial rights on paid plans.
- Watch out
- Hours caps, flat emotion, and cloning gated to higher tiers.
How to ask itvoiceId + style from the Voice Library + text with [pause 1s] tags + rate, pitch, variation 0 to 5, or audioDuration to fit a slotSpeech customization runs on Gen2 only, not the streaming Falcon model.
Use a pronunciation dictionary for brand names instead of respelling in the script.
Open Murf AI in Call Sheet →
6
Cartesia Sonic 3.6
Cartesia
Built for live voice agents. Fast and cheap, flat on long narration.
Very fast text to speech built for live voice agents. 44 languages, clones from a few seconds of audio.
- Cost
- Own platform, Pro US$5
- Where
- Cartesia's own site
- Rights
- Pro (US$5) carries a commercial use license.
- Watch out
- Tuned for phone‑agent speed, not performance.
How to ask itWell-punctuated full sentences + a voice that matches the language + optional speed, volume or emotion, or inline <break time="500ms"/> and <spell>ABC123</spell> tagsEnd every transcript with a full stop, question mark or exclamation mark.
Normal casing. All‑caps reads as an initialism, not shouting.
Open Cartesia Sonic 3.6 in Call Sheet →
7
Seed Audio 1.0
ByteDance Seed
Lines, other voices and the room around them in one pass, with each line's entrance timed in the prompt.
A whole sound scene from one prompt: speech and multi‑character dialogue, effects, ambience and room tone together, up to two minutes and continuable.
- Cost
- Not published on the maker's pages read
- Where
- KREA, Higgsfield · check in app: Magnific
- Rights
- Reference voices must be authorised.
- Watch out
- Timing control covers dialogue so far, not effects or music.
How to ask it[the speaker and their voice] + [the emotional tone] + [the line, and when it enters] + [the setting] + [the key sound effects] + [how the scene unfolds]Describe speaker, tone, delivery, setting, effects and how it unfolds.
Time the dialogue.
Open Seed Audio 1.0 in Call Sheet →