TTS Setup

Own speech setup, not script writing or a new synthesis framework. Use reasoning-first-narration for the explanation and existing specialist tools for synthesis and pronunciation.

Choose upfront

Inspect applicable accepted setups first: adopted media-use recipes, project production notes and any persistent user or environment memory available to the session, loading that memory if it is available but not yet loaded. Search the shared user or environment memory as well as the project's, because a voice the user accepted elsewhere is recorded there. A recorded listening-accepted default voice wins; use it unless the user asks for another. Installed voice or TTS skills, such as a personal voice-clone narration skill, are candidates only when no accepted default exists, and being installed is not acceptance. Carry forward their recorded resource context, credential references and account, tenant, region or provider restrictions, and verify that they still hold rather than rediscovering them. Reuse a listening-accepted choice without repeated onboarding, not merely an installed or operational one. If constraints exclude every accepted option, surface that conflict before proposing an unauditioned replacement. Otherwise ask once: local speech, a cloud service, or an existing setup? Human recording is also valid. A recommendation is not permission to install, spend, transfer private text or clone a voice.

For first setup, requested alternatives or a material failure, research current official docs, releases and model cards. Inspect actual OS/architecture, runtimes, CPU, available RAM/disk, GPU/backend and available memory, competing load, language, privacy/offline constraints, budget and duration needs. Do not select from a fixed catalog, personal default, GPU brand or parameter count.

Compare two or three suitable options, with dated source URLs, exact runtime/model/voice compatibility, rights/consent, maintenance, download burden, costs/privacy and uncertainties. Suitability includes audible quality for the intended delivery. Do not choose an installed system voice simply because it is available, local or free. Documentation benchmarks are not measurements on this host. Respect a requested cloud provider and explain text transfer, region/retention where known and likely charges. Let the user choose. Do not install or benchmark every candidate.

Set up only that path

Use the chosen official setup and existing specialist skill/adapter; inspect versions and current invocation schemas. Obtain needed mutation/service approvals. Isolate conflicting dependencies from the app and other media tools, preferably in a reusable user-scope environment. Verify selected assets, voice access, locale, license and official checksums where available. For cloud speech, verify the resource's owning account, tenant/subscription or project, and region as applicable. Use that explicit context rather than assuming CLI defaults.

Keep credentials in the selected integration's secure store or environment. Never print them or copy values into project files. Record only the non-secret invocation, runtime/model revision, voice/settings, rights and credential reference needed to reuse the choice, in the existing recipe, project production notes or the persistent memory entry the setup was reused from. Do not invent a profile database or elaborate schema.

Audition and reuse

Use a user-accepted spoken sample or an agreed delivery reference to establish the quality target. A reference sets naturalness and delivery expectations, not permission to clone its speaker. If none exists, establish the target through a short audition before full synthesis.

Synthesize a short representative passage with connected explanatory sentences, difficult product names, acronyms, identifiers, numbers and punctuation. Preserve canonical text separately from spoken adaptations. Decode and measure it, rejecting empty, invalid, clipped, truncated or silent audio. Then listen against the quality target for intelligibility, natural sentence stress, phrasing, pronunciation, pacing and sustained listenability. Reject robotic monotony, unnatural word spacing or distracting prosody even when every word is intelligible and measurements pass. A matching locale or voice name does not establish quality.

Record the quality target and intended delivery, exact sample, and who actually heard and accepted it. Only a human listener or audio-capable evaluator can supply that acceptance. Parent permission, waveform checks, an unread attachment, disclosure of missing listening or a promised later listen-through cannot substitute. Without listening acceptance, stop before full synthesis and obtain it.

Reuse audio only while text, voice, model and settings are unchanged. Recheck availability and relevant health, not the entire onboarding flow. Re-audition changed voices/models, difficult new terminology or quality problems. On failure report the observed stage (runtime, memory, voice access, credentials, quota, network or listening rejection), attempt bounded repair of the selected path, and preserve the last accepted sample. Do not silently switch engines, upload text, spend or install another model.