Our in-house voice model is the world’s best and the first built to adapt to video. It delivers your script in your voice, with your pacing and expression, while adapting to the scene’s acoustics.
Watch the launch video to see the #1 ranked voice model match its delivery to every setting.
HeyGen Voice reads a script the way you would, from a single 60-second recording. It keeps your timbre, accent, and how you lean on a word, then decides where the emphasis, the pauses, and the breath should go. Ranked #1 on the Artificial Analysis text-to-speech leaderboard and chosen over ElevenLabs v4 by 62% of listeners hearing their own voice.
1. Record your voice
2. HeyGen Voice
3. Speech for every scene
1. Record your voice
2. HeyGen Voice
HeyGen Voice
3. Speech for every scene
Expressiveness Boost is a separate layer trained on listener preference and scored against your reference recording the whole way through. One slider runs from your baseline delivery to full energy, and zero gives you the base model back.
Expressiveness Boost · Off
The baseline delivery
Four Similarity modes run from More adaptive, which tunes delivery to fit the script, to Closer to original, which faithfully matches your recorded take. Other models lock you into one fixed output, and HeyGen Voice lets you choose the mode for every line.
Similarity
← More adaptiveCloser to original →
Tap a mode to hear the same line. Play the original to compare.
One English recording reads Spanish, German, French, or Chinese with a native accent instead of a translated one. In cross-language tests, listeners picked HeyGen Voice as narrator over Cartesia Sonic 3.6 in 77.4% of decided votes.
Pick a language to hear the same voice. Play English to compare.
Cross-language narrator choicevs Cartesia Sonic 3.6
77.4%Share of decided votes · line = even split (50%)
Set a seed and the same line renders the same way across fifty takes or a 40-minute lesson. Instant Voices and Professional Voice Clones both hold speaker identity across narration, dialogue, and everything between.
Same line, same delivery, across every render.
HeyGen Voice is a standalone voice model. Generate speech from text or from your own recorded performance, download the audio, and use it anywhere: a podcast, an app, an ad, a course, or a voice agent. And when you make a video in HeyGen, the same voice is already there, matched to your avatar and tuned to the scene. Both paths run the same #1 ranked model.
Type a script, pick a voice or create your own, and hear human-like speech in seconds. The model reads emotional and contextual cues in the text and adapts its delivery to the sentence and the paragraph around it, so a question lands as a question and a list does not flatten.
“Wait, did you hear that? I think the launch is today!”
Record the line the way you want it said and HeyGen Voice returns it in a higher-quality voice, yours or another, keeping your emotion, emphasis, cadence, timing, and pacing. Direct the performance with your own voice instead of a tag vocabulary. Fix a flubbed take, swap the speaker, or lift a phone recording to studio quality. Available via the HeyGen API.
Your recorded takephone · 0:05
HeyGen Voicestudio quality · 0:05
Every voice model says it sounds the most realistic. We’d rather show you the data.
Read the technical reportIn the HeyGen app, thousands of people recorded a short voice sample. They heard the same script generated by HeyGen Voice, ElevenLabs v4, and Cartesia Sonic 3.6, unlabeled and in random order, and picked the one to use as their own.
See the dataChosen as their own voice
Three-way test · 6,153 votes
62.2%
Head-to-head vs ElevenLabs v4 · 8,332 votes
54.1%
Head-to-head vs Cartesia Sonic 3.6 · 7,286 votes
We also asked testers on Podonos which voices they would choose to narrate an entire lesson across 245 scripts in English, Spanish, German, French, and Chinese, 42,875 ratings in all. They chose HeyGen Voice over ElevenLabs v4 in 61.5% of decided votes, over Inworld TTS-2 in 68.1%, and over Cartesia Sonic 3.6 in 77.9%. It also led every system in clarity and native accent.
See the dataChosen to narrate a full lesson
HeyGen Voice’s share of decided votes
Share of decided narrator-choice votes, 245 cases, five languages, Holm-corrected p < 0.05.
A voice can sound exactly like you and still sound wrong for the shot. So we put the same voice in 50 places, including a podcast booth, cathedral, beach, and factory floor. We asked reviewers one blind question: does the sound match the picture? It did. Reviewers chose HeyGen Voice over Cartesia Sonic 3.6 in 73.3% of votes, over ElevenLabs v4 in 69.7%, and over Seedance 2.5 in 65.3%.
See the dataSound matches the picture
HeyGen Voice’s share of decided votes on ambient fit
Share of decided votes on ambient sound fit, 50 scenes × 5 reviewers, blind side-by-side video, Holm-corrected p < 0.05.
Listen to how the voice fits an open shoreline, with the sea stretching out behind the speaker.
Same script and reference voice within each scene. HeyGen Voice also receives the scene image.
Everything below is HeyGen Voice. Professional Voice Clone trains it deeper on you. Turbo serves it faster and cheaper for developers.
FreeUpload up to 60 seconds of audio from any device and HeyGen Voice keeps the person and drops the phone. Ready in minutes and works on HeyGen Voice and Turbo.
Spec
From $29/moGive HeyGen Voice 30 minutes to two hours of your recordings and it learns the finer details, like how your accent moves across words. It wins 63% of voice-similarity votes against ElevenLabs v4 and holds every inflection, pause, and quirk across a full course.
Spec
APIThe same model with a shorter generation step on dedicated hardware, for agents, narration pipelines, and anything at volume. Half the price of HeyGen Voice with lower latency, instant voices only, on the API.
Spec
“People have mentioned that it sounds much closer to my actual voice and a lot less robotic. Plus, it can finally pronounce my name correctly.”
Upload up to 60 seconds of audio from any device and HeyGen Voice keeps the person and drops the phone.
Without a recording, the model designs a voice for the person and place in the image.
Generated Speaker Caption“Mid-range, clear and articulate, upbeat and confident, moderate pace.”
Give HeyGen Voice 30 minutes to two hours of your recordings, and it learns the finer details, like how your accent moves across words.
Pitch variation, background level, and ambient steadiness are each a slider, so the same scene can play nearly clean or clearly busy.
Try it here
Freewith a HeyGen account
Four voices, twelve languages and up to 1,000 characters a take, right on this page.
Try it nowHeyGen app
Freein every video
Clone your own voice from a 60-second recording and use it in every HeyGen video. Professional Voice Clone from $29/mo.
Clone your voiceAPI
$30per 1M characters
Generate speech in any of your voices with one REST request. No packages, no minimums.
Open the API playgroundGenerate speech from your professional voice clone with one REST request, or stream audio as it’s generated.
REST and streaming endpoints with SDKs for [Python, Node, and others].
import osimport requestsresponse = requests.post( "https://api.heygen.com/v3/models/audio/tts", headers={"X-Api-Key": os.environ["HEYGEN_API_KEY"]}, json={ "voice_id": "your_professional_voice_id", "text": "Welcome back. Today we're recording from the studio.", "language": "en", "seed": 7, },)response.raise_for_status()audio_url = response.json()["data"]["audio_url"]with open("studio.wav", "wb") as f: f.write(requests.get(audio_url).content)
Keeps your voice
Set a seed to keep delivery consistent across renders