cartesia-sonic

What it does

The skill directs an agent to treat Cartesia Sonic as the primary low-latency text-to-speech provider for production needs, and to use adjacent Cartesia voice APIs only when explicitly required. It covers selecting sonic-3.5 as the default model, using various voice options by ID, applying pronunciation dictionaries, localization, voice cloning, and voice changing, and it provides production heuristics for when to prefer or avoid certain endpoints and workflows. It also documents endpoint behavior, recommended transcripts handling, and production defaults for audio output and generation settings.

How it works

• It specifies Sonic as the fast TTS model family with sonic-3.5 as the recommended starting point, noting latency and language support.
• It instructs to use different Cartesia surfaces (Sonic TTS, line-hosted agents, voice cloning, localization, infill, and voice changer) by exact surface names and not to conflate them as Sonic TTS.
• It outlines when to prefer Sonic (low time-to-first-byte, conversational pacing, multilingual needs, streaming) and when to use non-realtime or offline options.
• It requires recording endpoint, model ID, voice ID, language, output format, Cartesia-Version, generation_config, pronunciation dictionary ID, and transcript checksum in the asset manifest for pipelines.
• It details supported voices, cloning options (Instant Voice Clone, PVC), and guidance on testing voices for different production contexts.
• It enumerates endpoint choices for TTS requests (bytes, sse, websocket) with when each is best, and explains how to manage contexts and continuations.
• It lists core request fields (model_id, transcript, voice, output_format, language, pronunciation_dict_id, generation_config) and default production payload.
• It provides transcript and pronunciation guidance, including punctuation, codes, and digits handling, and warns about punctuation and casing effects.
• It includes pricing, quotas, and throughput planning (credits per character, cloning costs, concurrency limits, WebSocket limits, 429 handling), and the method to estimate credits.

When to use it

• Use as the production guideline whenever integrating, prompting, or reviewing Cartesia Sonic for production speech output. Use adjacent Cartesia voice APIs only if the production need explicitly requires them.
• Start with sonic-3.5 for new production; reserve other Sonic variants for regression or validated paths.
• Use WebSocket for live agents with streaming or incremental text; use Bytes or SSE when appropriate for transcripts with known timing but not WebSocket.
• Follow the asset-manifest recording requirement for pipeline reproducibility and auditing.

What it can touch

• Sonic TTS: model_id, transcript, voice, language, output_format, generation_config, pronunciation_dict_id, Cartesia-Version.
• Voices: voice IDs, Instant Voice Clone (playground/API), PVC (training states), voice localization, voice changer endpoints, and localization pricing.
• Endpoints: bytes, sse, websocket, and their context handling for transcripts and durations.
• Output formats: wav with pcm_s16le at 48 kHz as default, with alternatives like mp3 or raw PCM depending on needs.
• Pricing and quotas: character counts, clone costs, credits per second, concurrency limits, and 429 handling.

Caveats

• Model availability and pricing terms can change; the skill notes to re-check against Cartesia documentation as of 2026-07-10.
• Emotion controls are described as guidance rather than strict controls.
• PVC training has specific timing and cost considerations, including a 30-minute minimum and the need to retrain if a base model changes.
• Localization does not resolve translation or script timing; QA must cover pronunciation, intelligibility, and brand fit.
• WebSocket contexts expire 1 second after last audio output in some contexts; improper handling can lead to errors.