HeyGen's Voice Turbo Tops Speech Leaderboard at 87% Less Cost

HeyGen has replaced its own model atop Artificial Analysis’s Controlled Voice Arena. HeyGen Voice Turbo leads with a 1,202 Elo score, while Artificial Analysis estimates its price at $4 per million characters. That rate is 87% below the $30 estimate for HeyGen Voice, which now ranks second.

HeyGen Voice held first place for 12 days before Turbo overtook it. Both models were developed in-house by HeyGen, a company best known for generating avatar videos with synchronized speech. The new model is available through the API as heygen-voice-1-turbo.

Eight voices, blind votes

Artificial Analysis tests every model with the same eight cloned voices, reducing the influence of a listener’s preference for a particular voice. Participants hear two anonymous clips, choose the stronger result, and feed an Elo rating system borrowed from chess. Wins against highly rated models contribute more to the score.

Turbo’s score comes from 1,663 appearances. It ranks first for UK English and second for US English, behind Qwen-Audio-3.1-TTS-Plus. The six-point gap between the two HeyGen models is narrow, and Elo scores can move as more votes arrive.

A 7.5-fold price gap

Artificial Analysis lists Turbo as the least expensive model on the Controlled Voice leaderboard, despite its leading quality score.

Artificial Analysis derives both HeyGen figures from the company’s credit pricing. They are normalized estimates, rather than prices quoted directly by HeyGen, so production buyers should confirm their expected bill against current API terms.

HeyGen also offers heygen-voice-1-turbo-free through November 30, 2026. The free tier supports public voices and instant voice clones, requires no credit balance, and is documented for testing, prototyping, and limited production use.

Literal strings are Turbo’s strength

Artificial Analysis’s Pronunciation Robustness benchmark tests numbers, abbreviations, proper nouns, and other text that can expose failures hidden by general listening preferences. Human reviewers compare each clip with predetermined acceptable pronunciations across four categories. Turbo scores 87.8% overall, placing eighth among 33 models and improving on HeyGen Voice’s 83.2%.

Turbo’s first-place exact-sequence result matters for order numbers, product codes, account identifiers, and other strings that must be spoken verbatim. Abbreviation expansion remains weaker: Ming-3.0-tts leads that category with 96.5%.

Avatar economics explain the build

Low-cost speech generation directly reduces the expense of HeyGen’s core avatar-video product. Every generated presentation, training clip, or localized video requires synthesized audio, making per-character voice fees a recurring production cost.

HeyGen describes voice generation as part of the same stack that powers Avatar V, its core avatar model. A unified pipeline gives the company control over identity, speech, and synchronization while reducing its dependence on external voice providers such as ElevenLabs.

What ships in the API

Developers receive 10 shipped voices. The models API returns their voice IDs, which callers provide through the voice field of a text-to-speech request.

The HeyGen changelog documents POST /v3/models/audio/tts and a streaming variant. The default model remains heygen-voice-1, so callers must explicitly set heygen-voice-1-turbo to use Turbo.

• expressiveness_boost accepts values from 0.0 to 1.0 in instant mode.
• seed, speed, pitch_shift, pitch_variance, and break pauses require professional mode.
• OpenRouter also offers Turbo for teams consolidating model access and billing.

HeyGen markets Turbo as a faster model, although the cited quality and pronunciation benchmarks provide no latency measurements. Teams building conversational or real-time systems will need separate tests for time to first audio, streaming stability, and end-to-end response time.

One benchmark, one listening condition

The Controlled Voice Arena evaluates models while they clone the same reference voices. Its scores therefore describe cloned-voice output under a standardized listening condition.

Artificial Analysis’s separate Provider Voice Arena evaluates each vendor’s native shipped voices. Inworld appears on both boards, while HeyGen currently appears only on the controlled-voice board. Artificial Analysis omits models without enough votes, so that absence provides no basis for a quality comparison. Eleven v4 Turbo leads the native-voice board with 1,335 Elo.

Where the savings add up

High-volume workloads using a consistent cloned voice, including assistants, customer-service systems, documentation narration, and long-form audio, align closely with the benchmark’s test conditions. Turbo currently combines the board’s highest Elo score with its lowest listed price.

At the converted leaderboard rates, 10 million characters would cost about $40 with Turbo, compared with $300 for HeyGen Voice and $400 for Eleven v4 Turbo. Actual purchasing decisions still depend on voice fit, latency, reliability, cloning controls, and contract pricing, but Turbo materially changes the per-character calculation for large speech pipelines.