A short answer first: Devotel Orbit's stance on pre-tuned speech is that the default should be a reviewed decision and the override should be yours. The speech presets library that ships today on the Model Presets page has two layers. The curated layer is four named bundles, Balanced, High Intelligence, Ultra Fast, and Cost Saver, each pre-wiring the LLM, the transcriber, and the voice, and each carrying a curated latency, cost, and quality annotation. The override layer is the set of tenant-owned dials that sit on top of any preset: the per-agent speech-to-text engine catalog with its failover chain, the STT Playground that validates an engine on your audio before you switch, the per-agent voice and language settings, and the tenant-owned voice clone surface. This announcement walks what each STT and TTS preset stance covers, when to override versus stay preset, and how the governance underneath stays in your account.
1. Two layers, one library
The curated layer closes the recurring "which transcriber, which model, which voice" decision. Every preset bundles Deepgram as the transcriber and Cartesia as the voice, for one deliberate reason: a preset may only bundle engines that work for every tenant on the first call, without a bring-your-own provider key. The four presets differ in the reasoning tier and the latency lane, and each card shows the trade-off before you pick: Balanced runs Claude Sonnet 4.6 at roughly 900 ms to first audio byte and about $0.09 per minute; High Intelligence runs Claude Opus 4.7 at roughly 1400 ms and about $0.22; Ultra Fast runs Claude Haiku 4.5 on the low-latency lane at roughly 450 ms and about $0.05; Cost Saver runs Haiku off the lane at roughly 650 ms and about $0.03. Picking a preset instantiates a draft voice agent pre-wired with the bundle, and a wrong pick is cheap to reverse because the draft never takes traffic until you say so. How model presets trade latency against cost and quality walks the catalog card by card.
The override layer is where your market enters. The same engine catalog that powers the STT Playground drives per-agent speech-to-text selection and failover in production; the agent's voice and language settings stay editable after instantiate exactly like any other agent setting; and an ElevenLabs voice, including a cloned brand voice, is a per-agent upgrade you opt into. The presets keep the first call dependable; the dials keep every call after that yours. Read the library as six stances the two layers compose, and the next two sections define each stance against the surface that ships it.
2. The STT presets: accuracy-first, latency-first, telephony-codec-tuned
Accuracy-first is the stance where a misheard digit costs real money. On the curated layer it is the High Intelligence preset: the deepest reasoning tier, the lowest temperature, and the longest reply envelope, for collection negotiation and technical troubleshooting calls. On the override layer it is the engine you validated: run your recordings through the STT Playground, keep a human reference line, and if a catalogued engine beats the bundled default on your dialect mix, provision it under Settings → Voice → Speech-to-text and set it per agent with a failover chain behind it. Accuracy-first never means hand-tuning a model id; it means a reviewed preset plus, where your audio demands it, a validated engine swap.
Latency-first is the stance where hold time is the failure mode. On the curated layer it is Ultra Fast: the fast-model lane, roughly 450 ms to first audio byte, built for high-volume triage and short scripted flows. On the override layer it is bookkeeping: speech-to-text is the first slice of every agent's latency budget, and the per-engine response time the Playground reports is the number to carry into that budget before you commit a flow to the sub-second lane.
Telephony-codec-tuned is the stance the vendor clean test set hides. Cellular codecs, IVR digits, and crosstalk are exactly where meeting-room-trained engines drop, and no bundled default can be pre-tuned for every codec mix because your mix is yours. The library ships this stance as an instrument instead of a promise: record or upload a clip that is real for your business, and one Playground run sends the same bytes to every provisioned engine and returns each transcript, confidence score, and response time side by side. Deepgram stays the dependable out-of-the-box default; other catalogued engines sit labeled not yet available until you provision them, never as fabricated results. The STT Playground guide covers the evaluation workflow in full.
3. The TTS presets: natural-voice, voice-clone, low-latency-waveform
Natural-voice is the default register of the library: every preset ships a neutral Cartesia voice that reads as a competent agent on a phone call, not a demo. The starter voice is a known-good default, and the voice setting is yours after instantiate, so a team with a brand register picks its Cartesia voice per agent the same day it writes the prompt.
Voice-clone is the tenant-owned end of the range. The clone surface under Voice → Voice Clones trains a brand voice from consented recordings, with the owner consent captured before training, and applies it per agent as an ElevenLabs upgrade. It is deliberately not a preset default: a cloned voice carries a consent posture, and a consent posture belongs to the tenant that owns it, not to a bundle a platform ships. The voice clones explainer covers the custody model.
Low-latency-waveform is the voice side of the latency budget: how fast the first spoken byte lands after the model answers. Ultra Fast's roughly 450 ms number already includes it, which is the point of a bundle. The LLM tier, the latency lane, and the waveform are stamped together so the sub-second promise survives contact with a real call instead of being assembled dial by dial.
4. When to override, when to stay preset
Stay preset until evidence says otherwise. The four bundles exist so that a team of operators stops re-deciding the same three dials on every agent, and the metrics on each card, curated estimates of each tier's published profile rather than a per-call SLA, are enough to place most flows on the first pass. Override in one direction at a time, and only after measurement: the Playground for an engine swap, the voice setting for register, the language setting for a new market. The two how-to surfaces pair with this announcement on purpose: the model presets guide is the selection surface for the curated layer, and the STT Playground guide is the evaluation surface for the override layer. Announcements frame the stance; those two posts are where the clicks live.
5. Governance stays tenant-owned
Nothing in the library makes a compliance decision for you. Consumer disclosure sits in your account: the disclosure and consent prompts your agents speak are tenant-set per regime, and the presets inherit them rather than route around them. Latency budgets are tenant-set too: the time-to-first-token budget you configure per agent is the envelope a preset lands inside, re-checked when the agent runs, which is what makes Ultra Fast a reviewed placement instead of a vibes-based one. And the cost-per-minute envelope is visible before you commit: every preset card carries its estimated per-minute spend, and the same blended STT, LLM, and TTS decomposition shows up on your bill, so a Cost Saver rollout across a high-volume line is a number you approved, not a surprise you audited later. You set the posture; the platform enforces it at the decision point.
6. Three deployments, walked
English financial-services outbound. Dispute and payment-negotiation calls get High Intelligence, because the wrong answer is the failure mode that loses the customer, and the extra 500 ms earns more than it costs. The disclosure prompt your compliance team approved is set at the tenant, recording consent follows your posture, and the model dropdown is never hand-tuned: the preset is the reviewed decision, and the audit trail reads preset names, not pasted model ids.
Spanish healthcare inbound. This is the override layer doing its job. Instantiate Balanced for intake, then set the agent's language to Spanish and pick a voice that carries the register; run Spanish intake audio, ideally the code-switching caller, through the Playground and let the side-by-side decide whether the bundled default hears your patients or a provisioned engine does. Recordings and knowledge stay under the compliance posture you already run for health data. The preset supplies the reasoning tier; your market supplies the rest.
Multilingual retail support. Order-status and returns flows split by intent: Ultra Fast on the short scripted lane where responsiveness is the experience, Balanced where a return gets negotiative. Evaluate the language-switching clip in the Playground first, because a bundled default that transcribes a second language with first-language confidence is the silent failure in retail queues. One library, two layers, and the same tenant-owned governance row underneath all three.
Frequently asked questions
Are the six stances six more presets?
No. The catalog ships four presets; the stances are directions the two layers compose. Accuracy-first is High Intelligence plus, where validated, an engine swap. Telephony-codec-tuned is the Playground plus the per-agent engine catalog. Naming the stances keeps the library honest about what is curated and what is yours.
Why are Deepgram and Cartesia the bundled defaults?
Because they work out of the box without a bring-your-own key, which is the floor a one-click preset must hold. Per-agent upgrades onto other engines and voices stay available on the agent's own settings after instantiate.
Are the latency and cost numbers live measurements?
They are curated estimates reflecting each tier's published profile, a comparison aid at selection time, not a per-call SLA. Per-call truth comes from your agent analytics and your bill once the agent runs.
The takeaway
Pre-tuned speech is a reviewed default, not a locked pairing. Devotel Orbit's speech presets library ships four curated bundles with the transcriber and voice bundled for a dependable first call, and a tenant-owned override layer, engine catalog, Playground, voice and language settings, clone surface, for everything your market proves should differ. Stay preset until measurement says otherwise, override one dial at a time, and keep disclosure, latency budgets, and cost envelopes where they belong: in your account.
Published 15 October 2026.