Seed Audio Added to External TTS Engines
Set engine to tts-sa inside ExtParam and Seed Audio runs on the existing TTS APIs. With this addition, MPS now offers seven external TTS engines.
- Select the engine: pass
{"engine": "tts-sa"}inExtParam(a JSON string).synExt.modelcurrently supportstts-sa-v1only. - Select the generation mode: what you put into
engineCustom.referencesdecides the path — text only, text + image, or text + audio. - Select subtitles: set
engineCustom.audio_config.enable_subtitletotrueand word-level timestamps come back inExtInfo. The default isfalse.
It is the only engine that can generate speech from an image: pass a reference image in
references(image_urlorimage_data, at most 1 and ≤ 10 MB) and the engine picks up style cues from it.One request can cite up to 3 reference clips: address them as
@audio1/@audio2insidetext_prompt, where upload order maps strictly to the number. Getting the order wrong raises no error — it simply produces the wrong voice, so take care.Passing
referencesmakes the outerVoiceIdineffective. This is a parameter explicitly exempted from the "outer level wins for duplicate keys" rule; together with the identically named parameter ontts-fthere are only two.Cloning is not yet supported: there is no
CloneVoicepath, so no reusable voice. Reference audio does work, though.volis a[0.5, 2.0]multiplier, not the(0, 10]scale other engines use. Withogg_opus,sampleRatesupports48000only.It is slow, so the async API is recommended. Four APIs are supported:
TextToSpeech,TextToSpeechAsync,DescribeTextToSpeechAsyncTaskandDescribeVoices.