MPS TTS Engines Compared โ
The MPS text-to-speech APIs are engine-agnostic. Which engine actually runs the synthesis is decided by the engine field inside the ExtParam extension parameter. Seven external engines are currently available: ElevenLabs, Gemini, MiniMax, TTS-Seed, Fish Audio, IndexTTS and Seed Audio. They share the same APIs, authentication and billing, but their capability boundaries differ quite a bit.
This page puts all seven side by side to help you choose. Full parameters and request samples for each engine are in the per-engine docs linked at the bottom.
Pick one in a sentence โ
| What you need | Pick | Key capability |
|---|---|---|
| Karaoke-style subtitles, word-by-word A/V alignment | TTS-ElevenLabs | with_timestamps character-level timestamps |
| Describe accent and tone in plain language | TTS-Gemini | free-form engineCustom.prompt |
| Clone a voice, or design a new one from text | TTS-MiniMax | synthesis + cloning + voice design |
| News, ads, radio, ASMR and other broadcast styles | TTS-Seed | 24 emotion values |
| Custom emotion tags, or cloning without creating a voice | TTS-Fishaudio | inline tags in Text + zero-shot references |
| Use an audio clip as the voice directly, or sample voice and emotion separately | TTS-Index | reference audio URL straight in VoiceId + 8-dimension emotion vector |
| Generate speech from an image, or cite several reference clips in one request | TTS-Seedaudio | references takes images / up to 3 clips + @audioN addressing |
Capability matrix โ
| Capability | ElevenLabs | Gemini | MiniMax | TTS-Seed | Fish Audio | IndexTTS | Seed Audio |
|---|---|---|---|---|---|---|---|
engine value | tts-l1 | tts-gg | minimax | tts-s | tts-f | index-tts | tts-sa |
| Synthesis (sync / async) | โ / โ | โ / โ | โ / โ | โ / โ | โ / โ | โ / โ (async preferred) | โ / โ (async preferred) |
Voice cloning CloneVoice | โ | โ explicitly unsupported | โ | โ not yet supported | โ | โณ not listed (not needed) | โ explicitly "cloning not yet supported" |
| Zero-shot inline cloning (no voice to create) | โณ not listed | โณ not listed | โณ not listed | โณ not listed | โ
references | โ
the default mode (audio URL in VoiceId) | โ
reference audio via references |
| Speech from a reference image | โ | โ | โ | โ | โ | โ | โ
the only one (references.image_url) |
| Several reference clips addressable in text | โ | โ | โ | โ | โ | โ | โ
up to 3 + @audioN |
Voice design DesignVoiceAsync | โณ not listed | โณ not listed | โ | โณ not listed | โณ not listed | โณ not listed | โณ not listed |
| Emotion / style control | โณ not listed | โ
natural-language prompt | โ 7 enum values | โ 24 enum values | โ 24 inline tags, customisable | โ 5 paths (vector / audio / text / random) | โณ no dedicated parameter; via references and text_prompt |
| Voice and emotion sampled separately | โ | โ | โ | โ | โ | โ
emo_audio_url, independent of VoiceId | โ |
Volume vol | โณ not listed | โณ not listed | โ | โ | โ | โณ not listed | โ
[0.5, 2.0] multiplier |
Pitch pitch | โณ not listed | โณ not listed | โ
[-12, 12] | โณ not listed | โณ not listed | โณ not listed | โ
[-12, 12] |
Sampling controls (temperature / top_p) | โณ not listed | โณ not listed | โณ not listed | โณ not listed | โ | โณ not listed | โณ not listed |
| Quality / latency tier | โณ not listed | โณ not listed | โณ not listed | โณ not listed | โ
latency, 3 tiers | โณ not listed | โณ not listed |
| Timestamps | โ
with_timestamps (character-level) | โณ not listed | โณ not listed | โณ not listed | โณ not listed | โ overall duration only | โ
enable_subtitle (word-level, off by default) |
Model selection model | โ 6 options | โ 3 options | โ 2 options | โณ not listed | โ 3 options | โ 2 options | โ
tts-sa-v1 only |
Pass-through engineCustom | โ | โ | โ | โณ not listed | โ | โ 2.0+ models only | โ |
โ = explicitly supported by the integration guide; โ = explicitly unsupported; โณ = not mentioned in the integration guide, availability to be confirmed by an actual call.
โณ does not mean unsupported
A โณ only means "this engine's integration guide does not list it". The guides vary a lot in depth โ IndexTTS has the shortest one (3 APIs and a single normalized parameter), with TTS-Seed next, which is why those two columns carry the most โณ. If you actually need one of these capabilities, make a short test call first.
Only Seed Audio takes input beyond "text + audio"
For the first six engines the input is always "text (what to say) plus optional audio (whose voice to say it in)". The references array on tts-sa also accepts an image (image_url / image_data, at most 1 and โค 10 MB); the engine picks up style or scene cues from it while the words to speak stay in the text. The integration guide frames the three modes as text only, text + image, text + audio. It gives no mapping between image content and the resulting delivery, so test a few visually distinct images before putting this on a production path.
API coverage โ
All seven engines sit under the same MPS "Media AI โ Dubbing APIs" catalog, but each integration guide covers a different subset:
| API | Purpose | ElevenLabs | Gemini | MiniMax | TTS-Seed | Fish Audio | IndexTTS | Seed Audio |
|---|---|---|---|---|---|---|---|---|
TextToSpeech | Short-text synthesis (sync) | โ | โ | โ | โ | โ | โ | โ |
TextToSpeechAsync | Long-text synthesis (async) | โ | โ | โ | โ | โ | โ | โ |
DescribeTextToSpeechAsyncTask | Query async task | โ | โ | โ | โ | โ | โ | โ |
DescribeVoices | List available voices | โ | โ | โ | โ | โณ | โณ | โ |
CloneVoice | Voice cloning | โ | โ | โ | โ | โ | โณ | โ |
DesignVoiceAsync / DescribeDesignTask | Voice design and query | โณ | โณ | โ | โณ | โณ | โณ | โณ |
The common subset explicitly supported by all seven is the first three APIs: synthesis (sync / async) and task query. Code written against only these three can switch engines by changing a single engine value.
The fourth API is no longer part of the common subset
DescribeVoices used to be shared by all four engines, but neither the Fish Audio nor the IndexTTS integration guide lists it (Seed Audio brings it back, so five of the seven support it). If your code enumerates voices through DescribeVoices, that step may come back empty on tts-f or index-tts. Both have a fallback: tts-f can use the return value of CloneVoice or pass reference audio inline via references; index-tts goes further โ its VoiceId is the reference audio URL, so it needs no voice list at all.
IndexTTS has the narrowest subset, and that is enough
The index-tts guide lists only the first three APIs, not even CloneVoice โ the narrowest of the seven. That does not make it less capable: cloning is its default behaviour. Putting a reference audio URL into VoiceId achieves what other engines need a CloneVoice call for. What it lacks is the whole voice management layer, not the cloning itself.
Audio parameters โ
| Parameter | ElevenLabs | Gemini | MiniMax | TTS-Seed | Fish Audio | IndexTTS | Seed Audio |
|---|---|---|---|---|---|---|---|
Format format | wav (default) / mp3 / opus / alaw / ulaw | wav (default) / mp3 / pcm / ogg_opus / alaw / mulaw | wav (default) / mp3 / pcm / flac / pcmu_raw / pcmu_wav / opus | โณ not listed | wav (default) / mp3 / opus | โณ not listed | wav (default) / mp3 / pcm / ogg_opus |
Sample rate sampleRate | 8000 / 16000 / 22050 / 24000 / 32000 / 44100, default 16000 | 8000โ48000, default 16000 | same as ElevenLabs, default 16000 | same as ElevenLabs plus 48000, default 16000 | same as ElevenLabs, default 16000 | โณ not listed | 8000 / 16000 / 24000 / 32000 / 44100 / 48000, default 16000, no 22050 |
Speed speed | [0.7, 1.2] | โณ not listed | [0.5, 2] | [0.5, 2] | [0.5, 2] | โณ not listed | [0.5, 2] |
Volume vol | โณ not listed | โณ not listed | (0, 10] | (0, 10] | (0, 10] | โณ not listed | [0.5, 2.0] multiplier |
Target duration duration | โ | โณ not listed | โ | โ | โ | โณ not listed (output side returns ExtInfo.duration) | โณ not listed |
Five of them default to wav, not mp3
ElevenLabs, Gemini, MiniMax, Fish Audio and Seed Audio all return wav when format is omitted. For size-sensitive delivery (mobile downloads, for example) set mp3 explicitly. Note also that opus / ogg_opus / alaw / ulaw are not supported by the MPS built-in engine โ they only exist on the external engines.
This whole IndexTTS column is โณ โ do not borrow another engine's table
The index-tts integration guide gives exactly one synExt parameter, model. Format, sample rate, speed, volume and target duration are all unmentioned โ the only engine of the seven like this. If you need to control output format or speed, test those fields with short text first, and do not copy the tts-f or tts-s parameter table: that is another engine's declared scope and proves nothing about this one. The output side is explicit, though: ExtInfo.duration returns the total speech length in seconds.
Speed ranges, volume semantics and sample-rate options all differ
ElevenLabs only accepts [0.7, 1.2], while MiniMax, TTS-Seed, Fish Audio and Seed Audio accept [0.5, 2]. Volume is the subtler trap: vol is a (0, 10] scale on MiniMax / TTS-Seed / Fish Audio, but a [0.5, 2.0] multiplier on Seed Audio (0.5 = half volume, 2.0 = double) โ carrying vol: 5 over to tts-sa is out of range and fails validation. On sample rate, only Gemini (up to 48000), TTS-Seed and Seed Audio (48000 as an option) go above 44100; Fish Audio has no 48000 option, and conversely Seed Audio has no 22050. Reusing one parameter set across engines can fail validation for being out of range.
ogg_opus on Seed Audio comes with a hard constraint
With ogg_opus the sample rate must be 48000, while the default is 16000 โ writing only format: "ogg_opus" without touching sampleRate is a self-contradictory pair. Set both or neither.
Five schools of emotion and style โ
How you control delivery differs completely across the seven engines โ this is the difference most likely to bite you:
Enum-based (MiniMax / TTS-Seed) โ pick one value from a fixed list and put it in synExt.emotion. Controllable and reproducible, good for batch production.
- MiniMax, 7 values:
happy,sad,angry,fearful,disgusted,surprised,calm - TTS-Seed, 24 values: on top of basic emotions it adds scenario-style broadcast presets such as
news,advertising,storytelling,radioandASMR
Description-based (Gemini / IndexTTS) โ write one sentence of natural language. On Gemini it is engineCustom.prompt ("speak in a gentle whisper, slower than usual"); on IndexTTS it is engineCustom.emo_text ("You scared me to death! Are you a ghost?"), which the engine converts into an emotion vector. Most flexible, but the same sentence is not guaranteed to give an identical result every time.
Inline-tag-based (Fish Audio) โ emotion is not a parameter but part of the body: prepend a tag such as [excited] to Text. There are 24 common tags, and they are customisable rather than a closed enum.
Vector-based (IndexTTS) โ supply an 8-float array [happy, angry, sad, afraid, disgusted, melancholic, surprised, calm], each dimension continuously adjustable. This is the most precise and most reproducible option of the seven: enum-based control only picks discrete steps, while a vector can express "0.8 of sadness".
Not offered (ElevenLabs / Seed Audio) โ neither exposes a dedicated parameter such as emotion. On ElevenLabs style comes mainly from the voice you pick; on Seed Audio the style cues come from the references (audio / image) and text_prompt โ the only engine of the seven that substitutes reference material for emotion parameters.
IndexTTS spans two schools, and adds two more routes
index-tts offers both description-based (emo_text) and vector-based (emo_vector) control, plus two routes no other engine has:
- Emotion reference audio (
emo_audio_url+emo_alpha) โ take the emotion from a sample clip, i.e. "read it in this tone". - Random sampling (
use_random) โ ask for variety on purpose, e.g. generate several takes of one line and pick the best.
Five coexisting paths bring one problem: the integration guide gives no precedence between them. The safe approach is to use one at a time.
What makes inline tags unique, and what they cost
Inline tags are the only mechanism that supports changing emotion sentence by sentence within one piece of text โ enum-, description- and vector-based control all apply one baseline per request, so switching means splitting the request. The cost is that tags live inside the body: upstream copy assembly can easily end up reading the bracketed text aloud. Agree on a convention that emotion tags may only appear at the start of a sentence, and validate it before storage.
Division of labour
emotion sets the emotional baseline for the whole passage, while pitch / vol are physical adjustments layered on top of it. For a "low and calm" narration, emotion: calm + pitch: -3 (MiniMax) sounds more natural than tweaking pitch alone. Fish Audio has no pitch, but it does have temperature to control how far the delivery wanders โ lower it for stability. IndexTTS has neither, but emo_vector lets you push the "calm" dimension high directly, which amounts to setting the baseline. Seed Audio has pitch / vol but no baseline parameter โ to change the baseline you have to change the reference audio or the text_prompt.
Where VoiceId comes from โ
The first five engines take the engine's own voice name directly in VoiceId, and the naming styles differ widely. IndexTTS is entirely different: it takes an audio URL. Seed Audio depends on whether references is present.
| Engine | VoiceId example | Voice sources |
|---|---|---|
| ElevenLabs | ElevenLabs voice ID | system + cloned |
| Gemini | Kore (30 Gemini voice names) | system only |
| MiniMax | Chinese (Mandarin)_News_Anchor | system + cloned + designed |
| TTS-Seed | zh_female_vv_uranus_bigtts | system only |
| Fish Audio | 7f92f8afb8ec43bf81429cc1c9199cb1 | cloned + inline references audio |
| IndexTTS | https://example.com/voice.wav | the reference audio URL itself; no voice concept |
| Seed Audio | a system voice name; anything goes once references is set | system voices, or references (audio / image / speaker) |
Five engines are queried through DescribeVoices (ElevenLabs, Gemini, MiniMax, TTS-Seed and Seed Audio) with engine set in ExtParam. That API supports filtering by VoiceType (system / clone / design / all), Gender, Age, Languages.N and Labels.N, with pagination.
Fish Audio and IndexTTS are the two exceptions, and they are exceptions in different ways:
- The
tts-fintegration guide does not listDescribeVoices. Its voices come from the return value ofCloneVoice, or from dropping the reference audio intoengineCustom.references(at which point the outerVoiceIdhas no effect whatever you put in it). index-ttsgoes further: theVoiceIdfield is the reference audio URL. There is no "voice name" layer in between, and no extra parameter is needed to turn cloning on.
Choosing between the four cloning routes
Create a voice (ElevenLabs / MiniMax / Fish Audio): call CloneVoice once for a VoiceId, then reuse it. Cheaper on bandwidth when one voice gets heavy use. Inline references (Fish Audio): send the reference audio in engineCustom.references with every request, skipping voice creation and lifecycle management โ but you must supply the audio and its matching transcript. URL as voice (IndexTTS): write the audio URL straight into VoiceId โ no references wrapper, no paired transcript. The lowest integration cost of the four. References (Seed Audio): pass audio_url / audio_data in engineCustom.references; no paired transcript is needed, and up to 3 clips can travel in one request. It has no CloneVoice, so the result cannot be persisted as a reusable voice. The latter three all suit UGC (one voice per user, discarded after use) or batch jobs where voices far outnumber reuses, at the cost of re-fetching the audio on every request. IndexTTS adds one requirement: that URL must stay reachable long-term โ not a temporary signed link that expires.
Only IndexTTS can sample voice and emotion separately
On index-tts, VoiceId decides who is speaking and engineCustom.emo_audio_url decides with what emotion โ two independent clips. If you want "A's voice with B's emotion", it is the only one of the seven that can do it.
Only Seed Audio can switch voices within one request
Everywhere else, reference audio means "one voice for the whole request". The references array on tts-sa holds up to 3 clips, and text_prompt addresses them as @audio1 / @audio2 โ dialogue and multi-character reading no longer need separate requests. The catch: numbering is positional, not by name. The first array element is @audio1. Get the order wrong and nothing errors out; you simply get the wrong voice.
Migration notes and pitfalls โ
ExtParamis a string, not an object: the same for all seven engines โ it must be an escaped JSON string. Passing an object fails validation.- Outer level wins for duplicate keys:
synExt.xxxoverridessynExt.engineCustom.xxx. To use a vendor-only parameter, put it only inengineCustom, not in both layers. - The exceptions are the
referenceson Fish Audio and Seed Audio: pass it and the reference material determines the voice, overriding the outerVoiceIdโ the only parameter explicitly exempted from the "outer level wins" rule across all seven engines, and by coincidence both engines call itreferences. engineis matched exactly: no surrounding whitespace. The TTS-Seed and Seed Audio integration guides have both shown a leading space (" tts-s"," tts-sa") โ typos in both cases.- Check value ranges before switching engines: speed, sample rate and format options all differ; ElevenLabs' speed range is noticeably narrower, Fish Audio has no
48000sample rate, Seed Audio has no22050and treatsvolas a multiplier rather than a scale, and IndexTTS declares none of these parameters at all. - The meaning of
VoiceIdchanges: switching toindex-ttsturns it from a voice name into a reference audio URL; switching totts-sawithreferencesturns it into an ignored placeholder. A cross-engine wrapper must handle that layer explicitly. - Keep cloning and zero-shot apart: ElevenLabs, MiniMax and Fish Audio support
CloneVoice(a voice you can persist and reuse); IndexTTS does not listCloneVoicebut clones by default; Seed Audio states "cloning not yet supported" โ yet it does accept reference audio, which is close to zero-shot cloning in effect, except the audio must travel with every request and the result cannot be persisted. Gemini explicitly cannot clone, and TTS-Seed cannot yet. - Two engines now return timestamps, at different granularity: ElevenLabs'
with_timestampsis character-level; Seed Audio'saudio_config.enable_subtitleis word-level and off by default โ omit it and you get no timestamps and no error either. Word-level is usually the more useful one for Chinese. - The location of emotion control changes: on Fish Audio it moves out of
ExtParamand into theTextbody; on IndexTTS it moves back intoExtParam, but in the shape of a vector or a description; on Seed Audio there is no emotion parameter at all, only references andtext_prompt. A cross-engine wrapper has to absorb all of that. - IndexTTS'
emo_textis not the synthesis body: it only drives emotion inference; the words to speak stay inText. Mixing them up makes the engine read your emotion description aloud. - IndexTTS and Seed Audio both prefer the async API: both integration guides explicitly warn that they are slow โ the only two of the seven to say so.
- IndexTTS' emotion parameters are tied to a model version: the emotion parameters under
engineCustomare supported only byindex-tts-2(the default); explicitly selectingindex-tts-1.5makes them silently no-ops. - Seed Audio's
text_promptis not defined in the parameter table: all three generation modes are described in terms of it, yet the integration guide's parameter table never declares the field, and its relationship to the MPSTextfield is unspecified โ confirm it by testing before going live. - Gemini has a text length cap: text and prompt are each โค 4000 bytes, โค 8000 bytes combined. Split long text yourself.
- Fish Audio's default model has no published price: the price list covers only
s2-proands1(both $15.00 per million UTF-8 bytes), while omittingmodelselectstts-f-s2.1-pro. Confirm the billing basis before batch jobs.
Per-engine docs โ
| Doc | Engine value | Highlights |
|---|---|---|
| TTS-ElevenLabs | tts-l1 | character-level timestamps, 6 models |
| TTS-Gemini | tts-gg | prompt-driven style, no cloning |
| TTS-MiniMax | minimax | most complete: cloning + design |
| TTS-Seed | tts-s | 24 emotions, richest broadcast styles |
| TTS-Fishaudio | tts-f | customisable inline emotion tags, zero-shot cloning |
| TTS-Index | index-tts | reference audio URL in VoiceId, 8-dimension emotion vector, voice and emotion sampled separately |
| TTS-Seedaudio | tts-sa | the only engine that generates speech from an image, up to 3 reference clips addressable in text, word-level timestamps |
| AI Dubbing | โ | MPS dubbing overview and built-in engine |
| Media AI โ Dubbing APIs | โ | official API catalog |