HeyGen Voice Turbo vs Inworld Realtime TTS-2: The Best Cloned Voice Is Not the Fastest One

Ask both vendors for the number that decides a realtime product and only one of them answers. Inworld publishes a first-audio figure for Inworld Realtime TTS-2 — as low as 50 milliseconds on the Flash sibling, measured on Coval's board, with the vendor claiming up to 3× sooner first audio than ElevenLabs. HeyGen publishes no latency figure at all for HeyGen Voice Turbo; its documentation promises "faster generation" than the model it sits beside and leaves it there. What HeyGen does publish is a first place on the Artificial Analysis Controlled Voice Arena at 1,202 Elo and a listed $4.00 per million characters, against Inworld Realtime TTS-2's seventh place at 1,145 and $20.80. Voice-arena placings and per-character prices that move after publication are recorded with their dates and sources in the AI Incident Archive, alongside the other model and vendor incidents we follow.

So the two models are not competing for the same slot. One is the best cloned-voice engine on a blind-preference board with an unanswered latency question. The other is a realtime engine with a published latency claim and a quality score five places down. Which one you want depends on whether a human is waiting mid-sentence.

The two boards, and where each vendor actually stands

Artificial Analysis runs both speech boards, and the pair of numbers matters more than either alone because they measure different things: the Controlled Voice Arena holds eight cloned reference voices constant and scores the engine; the Provider Voice Arena lets each vendor speak in its own native voices and scores the shipping product.

• Controlled Voice Elo — HeyGen Voice Turbo 1,202 (#1 of 46, ±19, 1,685 samples). Inworld Realtime TTS-2 1,145 (#7, ±13, 3,161 samples).

• Provider Voice Elo — HeyGen Voice Turbo not listed. Inworld Realtime TTS-2 1,267 (#6 of 98).

• Listed price per 1M characters, arena conversion — HeyGen Voice Turbo $4.00. Inworld Realtime TTS-2 $20.80, Flash $10.40.

• Vendor rate card — HeyGen publishes no per-character voice rate; the $4.00 is derived from credit pricing. Inworld publishes $25.00 on demand for Realtime TTS-2 and $15.00 for Flash, dropping to $20.00 and $10.00 on the Creator plan.

• Stated latency — HeyGen: none published. Inworld: 50 ms to first audio for Flash on Coval; the flagship's figure is not stated on the same basis.

• Language coverage — HeyGen: 42 base language codes for cloning and speech. Inworld: about 500 languages and dialects for one Realtime TTS-2 voice.

The asymmetry in the Provider column is the one to notice. Inworld is ranked on both boards; HeyGen is ranked on one. Artificial Analysis omits models that lack enough votes, so that absence is not a quality signal — but it does mean that Inworld's product has been scored across its whole shipping surface and HeyGen's has been scored only in the cloned-voice condition.

Quality: a 57-point gap that survives the error bars

Fifty-seven Elo is the largest gap in this comparison and the least ambiguous. HeyGen Voice Turbo's interval runs 1,183 to 1,221; Inworld Realtime TTS-2's runs 1,132 to 1,158. They do not overlap. Unlike the seven-point HeyGen-versus-HeyGen comparison, where the two models are statistically tied, this one has separation at 95% confidence.

What that buys, precisely: on eight cloned reference voices, blind listeners preferred HeyGen Voice Turbo's renditions by a margin the board can resolve. It does not say HeyGen's own voice catalogue sounds better than Inworld's — HeyGen has no Provider board entry to make that claim with — and it does not say Turbo clones better, only that it speaks a supplied clone better.

Inworld's cloning path is the one its quality score is least able to describe. Instant cloning works from three seconds of reference audio, and voice design works from a written description; both are included on every plan. The arena measures the speech that comes out, not the fidelity of the clone that went in. A model that scores 1,145 on a clone it was handed may still be the better choice for producing that clone from three seconds of a customer's voice.

Latency: the number HeyGen will not give you

Inworld's realtime positioning is specific and falsifiable. Realtime TTS-2 Flash returns first audio in as little as 50 ms on Coval's measurement, up to 3× sooner than ElevenLabs at 183 ms and up to 35× sooner than O​penAI's GPT-4o mini T​TS at 2,056 ms. Those are third-party measurements of a vendor's model, quoted by the vendor — better evidence than a marketing claim, weaker than an independent audit, and worth reproducing on your own network before you architect around them.

HeyGen's documentation describes Turbo as speaking instant voices "at lower latency" than the default and stops there. Artificial Analysis's throughput page lists a median generation time of 7.29 seconds and 122 characters per second for the model, with a 5th percentile of 52.7 seconds and 95th percentile of 150 seconds — figures that describe whole-request synthesis on its measurement rig, not time to first audio, and that variance is wide. A seven-second median is fine for pre-rendered narration. It is not a conversation.

That comparison is not a fair fight, and it is not meant to be. HeyGen's Turbo is priced and documented for volume synthesis: 5,000 characters per request, 30 requests per minute, a synchronous WAV endpoint and an SSE stream whose parts are complete WAV containers. Inworld's is documented for turn-taking: instruction-based delivery control, 50 non-verbal sounds rendered as real audio, and a first-audio claim in tens of milliseconds. If your product speaks inside a conversation, HeyGen has not published the number you need, and absence of a number is not evidence of a slow model — it is evidence that this is not the product they are selling you.

Price: three rungs on one side, one derived figure on the other

The $4.00 against $20.80 looks like a five-to-one win for HeyGen, and it is the least reliable number in this article. It is Artificial Analysis's conversion of HeyGen's credit pricing, derived because HeyGen does not publish what a credit buys in characters. Inworld's $20.80 is the same board's normalisation of a rate Inworld does publish.

Compare like with like and the picture narrows. Inworld's own ladder runs $25.00 on demand, $20.00 on the Creator plan, $17.50 on Pro, $15.00 on Business and $12.50 on Enterprise, with Flash at $15.00 down to $7.00 across the same tiers, and the vendor advertising sub-$5 at the top. At the rates a committed customer actually pays, Realtime TTS-2 lands between $12.50 and $20.00, and Flash between $7.00 and $15.00. Against a $4.00 conversion, HeyGen is still cheaper on paper — but the gap is closer to two-to-one against a discounted Inworld tier than five-to-one against list.

There is a second rung that costs nothing. HeyGen's heygen-voice-1-turbo-free tier speaks public voices and instant clones without a credit balance through 30 November 2026, on shared capacity with no guarantees and a documented tendency to return 429 under load. Inworld offers up to 70 minutes of free T​TS on its On-Demand plan. Both are enough to run a real comparison; neither is enough to run a product.

Where each model stops being the answer

Reach for HeyGen Voice Turbo when the audio is pre-rendered or near-real-time rather than conversational: narration, audiobook chapters, e-learning modules, the spoken half of a video pipeline, or a voice agent whose turns are long enough that seven seconds of synthesis hides behind a thinking indicator. Its case is a first-place cloned-voice score at the lowest derived rate on the board, with 42 languages and a 44.1 kHz WAV out.

Reach for Inworld Realtime TTS-2 when a person is waiting. The latency claim, the instruction steering — send delivery direction as a request parameter or in brackets mid-line — the 50 non-verbals, the ~500 languages and the three-second clone are all features of an engine built for dialogue, and no amount of Elo closes that structural gap. Its case also includes a Provider board rank HeyGen does not have, which matters if your product ships stock voices instead of clones.

Neither model is the answer if you need a professional clone with reproducible seeds: HeyGen restricts Turbo to public and instant voices, and prosody controls like seed, speed and pitch_shift are professional-only on that API. Inworld's equivalent controls are its own instruction and steering layer, which is a different mechanism with different guarantees.

Running both without betting on either

The awkward part of this decision is that the two candidates optimise for conditions you cannot fully model from a leaderboard: your network's time to first audio, your text's names and numbers, your listeners' tolerance for a pause. The cheap way to answer it is to put both behind one interface and split traffic — a small share to the challenger, automatic failover to the incumbent when a request fails or a rate limit fires.

That is the shape OrcaRouter's routing layer is built for, and it is the honest reason to mention it here: one key for 200+ models, provider list price passed through with no markup, and failover that turns a 429 from a shared-capacity free tier into a retry rather than a dead air moment. HeyGen Voice Turbo is not in that catalogue today, so a genuine HeyGen-versus-Inworld split still means two integrations — but the failover half of the argument holds regardless of which model you end up standardising on.

What the boards settle is narrower than they appear to. HeyGen Voice Turbo is the better cloned-voice engine by a margin the statistics support, at a price that is probably the lowest in the top tier. Inworld Realtime TTS-2 is the only one of the two that has told you how fast it starts speaking, and for a conversational product that single unanswered question outranks fifty-seven Elo.