Where the voice comes from, model by model

Eight entries generate the sound in the same pass as the picture. Six are handed a track, or a script to read out, and asked to make a face agree with it. Two of the 16 models publish nothing about where the sound comes from. One publishes a language catalogue with codes; the rest publish a number or nothing. From 11 source pages, read 2026-09-22 to 2026-10-25; no maker pays to be listed.

Inclusion rule. Products whose own English documentation says something about video that carries speech, whether the sound is generated with the picture or supplied to it. A field the vendor does not document is recorded as not documented rather than inferred from generated output. Order. Alphabetical by model name.

1Three architectures sit in one table, and they fail differently

Eight products here make the sound while they make the picture, so alignment arrives for free and the words are the part that resists direction. Seven take a finished recording or a script and animate a face to it, so the words are exact and the picture has to follow. Two document neither, which leaves a buyer unable to tell from the vendor which of the two shapes is being bought.

That difference decides a schedule rather than a preference. A generated-audio route can be prompted, heard and revised inside one pass. A supplied-audio route needs casting, a recording session and a delivered track before a single frame exists, and it gives a director something to approve before any rendering is paid for.

2Keeping a voice: a stored object, an id on the call, or nothing

Five entries let a voice exist before a shot does, bound to an element, attached to a character, matched from a library with a sample to play, or created as a named object with a preview to approve first. Five re-establish it on every call, from a reference clip capped at fifteen seconds or a speaker id chosen per request. The remainder publish nothing about persistence, which across forty episodes amounts to a warning.

Both documented arrangements produce a consistent voice, and they break differently. A stored voice is wrong once if it is wrong at all. A per-call clip is an opportunity to drift on every generation, and the drift is silent: nothing in a returned file records which sample produced the voice in it.

3Languages get counted far more often than they get listed

One vendor publishes rows, each voice carrying a formal language name, a native name and a language code. One names fifteen dialogue languages and adds Cantonese for speech alone. The rest offer a number without names, a bare language field, or silence. The largest figure here, ninety-five or more, belongs to a model that works from a waveform and so has no per-language voice inventory to publish in the first place.

A count and a list are different assets. A count says the capability exists somewhere inside the product; a list says whether the market a producer was asked about is inside it, which is the question that gets asked before a season is commissioned rather than after.

4Two speakers in one frame, addressed once

The commonest shot in drama went unaddressed by every vendor in the first reading of this register. One now addresses it: a prompting guide asks for speakers to be labelled consistently and turns alternated, so each line lands on the right face. Another ties lip movement to whoever is on screen, which implies a selection is being made without describing how. A third warns that lip sync works better framed closer, which is a quiet way of saying the wide two-shot is the weak case.

Three sentences, then, on the shot a dialogue scene is built out of. Everything else in the register is silent about it, and silence here is not evidence that the case is handled.

5Per model

• D-ID — audio source: a script or an audio url; voice per character: a voice id from five providers.
• HeyGen — audio source: a separate speech endpoint; voice per character: a clone reused by id.
• Kling AI — audio source: with the picture, VIDEO 3.0; voice per character: bound to elements.
• LTX Studio — audio source: with the picture, plus audio-to-video; voice per character: on character Elements.
• Luma Ray — audio source: not documented; voice per character: not documented.
• MiniMax — audio source: with the picture; voice per character: reference audio, 15 s per call.
• PixVerse — audio source: uploaded audio, or text read by a voice; voice per character: a speaker id from the voice list.
• Runway — audio source: not documented; voice per character: a stored voice with an id.
• SceneMixer — audio source: with the picture, in the chosen language; voice per character: a matched preset, kept on the character.
• Sora 2 — audio source: with the picture; voice per character: speaker labels in the prompt.
• sync-3 — audio source: supplied or read from text; voice per character: not documented.
• Synthesia — audio source: from the script, or uploaded; voice per character: catalogue or a cloned voice.
• Veo — audio source: with the picture; voice per character: not documented.
• Vidu — audio source: with the picture, speech-only available; voice per character: not documented.
• Wan 3.0 — audio source: on by default; voice per character: reference audio, 15 s per call.
• Wan2.2-S2V — audio source: supplied to the model; voice per character: the supplied track.

• Audio source
  Described as a media generation model generating videos with synced audio, with video and audio both listed as outputgenerated with the pictureOpenAI, Sora 2 model page / recorded 2026-09-22
• Two characters in frame
  For multi-character scenes the guide asks for speakers labelled consistently and turns alternated, so each line lands on the right characteraddressed rather than skippedOpenAI, Sora 2 prompting guide / recorded 2026-09-22
• Voice source
  A custom voice can be built from an audio sample between 10 seconds and 5 minutes long and at most 10 MBa published sample windowRunway, custom voices reference / recorded 2026-09-22
• What choosing silence costs
  Enabling or disabling audio does not affect pricingAlibaba Cloud, Wan 3.0 API reference / recorded 2026-09-22
• Sound
  Plans do integrate the following audio models into the Luma generative ecosystem: ElevenLabs SFX, Music and v3named as an integration to comeLuma, information for AI assistants / recorded 2026-09-22
• Voice source
  A voice clone is instant from one recording or professional from twenty minutes or more, and is then passed as a voice idtwo enrolment routes with a thresholdHeyGen, API quick start / recorded 2026-09-22
• Languages
  Each voice is listed with its formal and native language name, a language code, a gender, a name and a voice idpublished as a catalogue with codesSynthesia, list of supported voices / recorded 2026-09-22
• Voice source
  The speaker list returns the built-in voices and the account's custom voices, and can be filtered to either kindcatalogue or supplied samplePixVerse, Get TTS Speaker List reference / recorded 2026-10-25
• Languages
  sync-3 supports 95 or more languages, described as the same coverage as the models before itpublished as a countSync, sync-3 model documentation / recorded 2026-09-22
• Voice source
  Five speech providers are named for the voice: Microsoft, ElevenLabs, Amazon, Google and Azure OpenAIoutside catalogues named on the pageD-ID, create a talk reference / recorded 2026-09-22
• Audio source
  The card describes audio-driven cinematic video generation from an audio input with a reference image and an optional text prompthanded to the model rather than made by itWan-AI, Wan2.2-S2V-14B model card / recorded 2026-09-22

6Sources

Each row links the page it was read from, and each entry carries its own reading date. What counts as documented is on how read; the fields themselves are described on the field notes.