TTS Director

Use this skill before calling Operate(resource="voice", operation="speech", input={ ... }) when the speech should sound performed rather than flat.

Core Mission

Turn the final spoken script into one executable TTS request:

• Preserve the user's words unless they asked for rewriting.
• For dialogue or role audio, first identify the speaker/role count and choose a distinct voiceId for each role before segmenting. Use availableVoicesForAgent from the generation context when present; call voice.list if the context does not provide enough voices.
• Identify the selected TTS model before designing controls. Use model from the request/context when present.
• For CosyVoice-family models such as cosyvoice-v3.5-plus, only invoke cosyvoice-ssml when the current task is already a video-director managed digital-human / VideoRetalk / asset-library talking-head video and the TTS is the approved character speech track. Outside that narrow flow, do not activate cosyvoice-ssml.
• For MiniMax-family models such as speech-2.8-turbo, split the script into meaningful performance beats and speaker turns, then assign each beat emotion, speed, pitch, punctuation, and pause markers.
• Submit one voice.speech call. Do not synthesize each segment manually.
• In the final user-facing response, present the merged finalAudio only. Do not list segment files as the result.

Model Branching

Before creating the final payload, classify the TTS model:

• CosyVoice branch: model name contains cosyvoice, especially cosyvoice-v3.5-plus; this branch may use cosyvoice-ssml only inside a digital-human / VideoRetalk video flow already owned by video-director.
• MiniMax branch: model name starts with speech- or contains minimax.
• Unknown model: keep controls conservative; prefer plain input plus prompt only if the provider is known to accept it.

This distinction is mandatory because the models support different controls: