AI voice products increasingly arrive as integrated studios, agents, and localization systems. That makes their labels deceptively simple. “Text-to-speech,” “speech-to-speech,” and “voice cloning” are often presented as competing choices, although they describe different parts of a system.
Text-to-speech and speech-to-speech describe the media entering and leaving a pipeline. Voice cloning describes how a speaker identity is created or represented. A cloned voice can drive text-to-speech. It can also condition a speech-to-speech transformation. A product may combine all three and expose only one button.
The distinction matters in 2026 because recent launches span very different jobs. xAI’s Grok Voice Agent Builder targets live, tool-using conversations. ElevenLabs’ Dubbing v2 targets localized performances that remain close to original delivery. Both involve audio in and audio out, but their latency, editing, review, and success criteria are not interchangeable.
The shortest useful definition of each term
| Technology | Primary input | Primary output | What it controls best |
|---|---|---|---|
| Text-to-speech (TTS) | Written text | Generated speech | Wording, pronunciation, and repeatable narration |
| Speech-to-speech (S2S) | Spoken audio | Spoken audio | Conversational response or transfer of an existing performance |
| Voice cloning | Reference recordings or a voice profile | Reusable speaker conditioning | Vocal identity and similarity |
The first two rows are transformations. The third is a capability that can sit inside either transformation. Keeping those layers separate makes product comparisons more honest and system design more flexible.
TTS: the script is the source of truth
In a TTS workflow, text determines what is said. A model converts that text into speech using a selected stock voice, designed voice, or consented clone. The source may include pronunciation hints, language, emotional direction, or timing controls, but the core contract remains text in and audio out.
TTS is a strong fit for narration, accessibility audio, product explainers, frequently updated lessons, and generated dialogue whose wording must match an approved script. It is easy to rerun one sentence after a copy change, and text can pass through legal or editorial review before synthesis.
Its weakness is that performance must be specified or inferred. A written line does not fully encode breath, hesitation, emphasis, tempo, or the relationship between speakers. Advanced models can produce expressive speech, but a prompt asking for “warm urgency” is not equivalent to an actor’s actual delivery.
TTS quality should therefore be evaluated on more than naturalness. Test pronunciation of names and domain terms, consistency across paragraphs, pacing, language switching, numerical readings, audio artifacts, and repeatability. If a cloned voice is used, add similarity and unauthorized-use controls to the evaluation.
Speech-to-speech has two important meanings
Speech-to-speech begins with audio, yet that phrase now covers at least two product families.
The first is conversational S2S. A person speaks, the system interprets intent, reasons or retrieves information, may call tools, and responds in speech. xAI describes Grok Voice Agent Builder as an end-to-end speech-to-speech path rather than a conventional stack assembled from separate speech recognition, language-model, and TTS services. Its July 2026 beta combines telephony, retrieval, tools, guardrails, MCP, and observability. xAI reports support for more than 25 languages and publishes its own voice-agent benchmark results; those are vendor claims that should be tested on the accents, interruptions, tools, and call conditions of a real deployment.
The second is performance-conditioned transformation. ElevenLabs says Dubbing v2 conditions on the original performance and transfers aspects such as tone, pacing, delivery, and emotion across more than 90 languages. Its sync-aware translation is intended to keep translated speech aligned with source timing. That is valuable for film, education, and creator localization, but translation accuracy and cultural suitability still need human review. The launch announcement made Dubbing v2 available in ElevenCreative and ElevenProductions while describing an API as forthcoming, so teams should verify current API availability rather than assume the product interface and API shipped together.
A live support agent and a dubbed documentary are therefore both S2S workloads, but one optimizes turn latency, interruption handling, and correct tool calls. The other optimizes semantic fidelity, performance continuity, timing, and mix quality.
Voice cloning is identity, not a complete workflow
Voice cloning derives a speaker representation from reference audio. That representation may be used to read arbitrary text, transform another performance, or maintain one brand voice across languages. The clone does not decide the script, translate it, remove background noise, synchronize a face, or publish the result unless the surrounding product adds those stages.
Reference requirements vary. xAI says its Voice Agent Builder can create a brand voice from about two minutes of audio, alongside built-in voices. That is a product-specific statement, not a universal minimum for high-fidelity cloning. Recording quality, speaking range, language, emotion, and model design all affect results.
Cloning also changes the risk profile. A generic narrator voice can still mislead, but a recognizable clone can enable impersonation and reputational harm. Obtain explicit permission, document permitted uses, restrict access to reference recordings and voice IDs, provide revocation, and disclose synthetic speech where appropriate. For public figures, employees, customers, and minors, policy and legal review should precede generation.
How to choose the right path
Start with the artifact and its acceptance test rather than a model label.
- Choose TTS when an approved script must be rendered reliably and revised sentence by sentence.
- Choose performance-conditioned S2S when preserving delivery from an original recording matters more than generating from a blank script.
- Choose conversational S2S when the system must listen, reason, use tools, and respond with low perceived latency.
- Add voice cloning only when a specific identity is necessary and authorized; otherwise a licensed stock voice is simpler.
- Combine approaches when localization needs a reviewed translation, an authorized identity, and timing close to a source performance.
Run evaluation with representative inputs. For a voice agent, measure time to first response, interruption recovery, transcription errors, tool-call success, escalation to a human, and call completion. For dubbing, use bilingual reviewers to check meaning, names, timing, emotion, and speaker attribution. For TTS, review long-form consistency and corrections. Never infer production quality from one polished demo.
Evaluate operational fit as well. Record whether the product exposes editable transcripts, pronunciation controls, speaker separation, stems, timestamps, version history, and export formats the downstream editor can use. Check how failed generations are billed, how reference recordings can be deleted, and whether a voice profile can be restricted to one workspace. A technically convincing sample can still be unsuitable when identity governance or revision handling is weak.
Where a composable Medux workflow fits
An integrated voice-agent or avatar product can reduce setup, while a composable workflow gives teams explicit control over individual media stages. Medux belongs in the latter category in the linked tutorials. It is not described there as a native component of Grok Voice Agent Builder or ElevenLabs Dubbing v2, and connecting it through Codex or Claude requires separate MCP configuration and authorization.
The Medux voice cloning and TTS tutorial demonstrates a concrete asynchronous path: Codex reads an approved text file, requests an upload location for reference audio, creates a cloning task, polls its status, sends the text for synthesis, and downloads a WAV result. This can serve script-led narration where the voice owner has consented. It should not be presented as live speech-to-speech or telephony.
After synthesis, a producer can treat audio finishing as another bounded operation. The Claude pitch-adjustment tutorial creates higher- and lower-pitch variants as separate jobs and checks each task ID before downloading the MP3 files. Pitch shifting is a creative transform, not a substitute for a properly authorized clone, and every version still needs listening review.
For a talking-head deliverable, the Codex lip-sync tutorial uses a source avatar clip plus finished audio, uploads both assets, starts an audio-driven task, polls until rendering completes, and downloads an MP4. That downstream step is distinct from generating or cloning the voice.
A reliable sequence is therefore: approve the script and identity rights, generate or transform speech, validate pronunciation and timing, optionally adjust the audio, then render and review lip sync. Preserve the task ID for each stage and never resubmit a completed stage merely because an agent session reconnects. The layered design makes failures easier to diagnose and lets a team replace one voice model without rebuilding every media operation.