# Choose your AI models use by use, based on measured trials

> Voice, transcription, vision, analysis: each use has its own provider and model. The Voice Lab compares them in real conversations before you choose.

Source : https://tellmehow.io/en/features/provider-agnostic-ai-and-voice-lab · Updated on : 2026-10-18

Voice, transcription, vision, analysis, chat, and search are configured separately. The Voice lab lets you talk with each setup and measures latency and cost before you use it on an engagement.

The TellMeHow AI layer is provider-agnostic: each of its six uses (voice, transcription, vision, analysis, chat, and search within the audit) is configured with its own provider, model, and settings, at the firm level, then per method and per engagement.

- 6 AI uses, each configured separately
- Two voice architectures: direct speech-to-speech, or transcription, model, synthesis
- Firm API keys encrypted and never displayed

### A tool tied to a single model

Voice and language models change every quarter, and so do their prices. A tool that imposes its provider also imposes that provider’s quality in each language, its latency, where data is processed, and its bill. The firm has no way to compare, let alone switch.

### Six uses, six configurations

Running the conversation, transcribing, reading the screen, analyzing, answering in the chat, and searching the audit have different constraints. Voice needs low latency, analysis needs reasoning over long context, and vision needs a controlled cost per image. So each use gets its own provider, model, and settings.

- Profiles defined for the workspace, overridable per method and then per engagement.
- Default profiles provided, all of them editable.
- A “test” button on every profile that shows the trial’s latency and cost.
- Vision kept separate from voice by default, so every screen reading is stored and verifiable.

### Two ways to give the AI interviewer a voice

With direct speech-to-speech, a single model listens and talks: latency is lowest and interruptions are handled natively. With a pipeline, speech is transcribed, a language model of your choice replies, and speech synthesis reads it out: every component is swappable, at the cost of slightly higher latency. The rest of the product does not see the difference.

- Direct voice: OpenAI Realtime.
- Pipeline: OpenAI or Deepgram transcription, an OpenAI or OpenAI-compatible language model, OpenAI or ElevenLabs synthesis.
- Either way, the AI interviewer has the same tools: topic coverage, screen reading, and pins.

### The Voice lab, for evidence-based choices

The Voice lab assembles any configuration: direct voice or pipeline, provider, model, and voice for each component, language, with or without screen sharing. You talk with it for three minutes, using the guide from one of your engagements. Time to first audio, average response delay, and cost are measured and stored in the trial history.

- A sortable history that flags the best result and keeps your notes on each trial.
- A configuration you already tried can be rerun as is for comparison.
- Trials are shown separately from engagements in cost tracking.

### Keys, security, and sovereignty

The firm uses its own API keys, verified, encrypted, and never displayed, or the platform’s keys if its plan includes them. For analysis, chat, and search, it can choose European options: Mistral, or self-hosted models behind an OpenAI-compatible endpoint. Screen text recognition, which runs before redaction, happens on our own servers.

- What the participant says and what appears on screen are treated as data, never as instructions.
- The compliance kits list the providers actually configured as subprocessors.

### In practice

- **Uses** : Voice, transcription, vision, analysis, chat, and search within the audit; the voice pipeline adds a live transcription profile and a speech synthesis profile.
- **Providers** : OpenAI (default), Mistral, Deepgram, ElevenLabs, and any OpenAI-compatible service (OpenRouter, Ollama, vLLM, and others), depending on the use.
- **Text recognition (OCR)** : An engine running on our servers for redaction; for artifact text, the engine of your choice, such as Mistral OCR.
- **Configuration levels** : Workspace, then method, then engagement; the screen shows the effective profile for each use and where it comes from.
- **API keys** : Verified online before saving, encrypted at rest, never displayed or logged.
- **Voice lab** : Three-minute trials in real conversation, reserved for workspace admins, with latency and cost measured.
- **Session languages** : English and French.
### Frequently asked questions

**Can we use our own API keys?**

Yes. A firm stores its own keys in TellMeHow, provider by provider: they are verified, encrypted, and never displayed again. It can also use the platform’s keys if its plan includes them. Cost tracking shows which key was used for each call.

**What is the difference between direct voice and a pipeline?**

With direct voice, a single model listens and answers out loud, which gives the lowest latency and native handling of interruptions. With a pipeline, transcription, language model, and speech synthesis are three components chosen separately, so you can pair the best voice for a language with the model of your choice. TellMeHow supports both, and the Voice lab lets you compare them.

**Why is vision separate from voice by default?**

Screen-reading errors have been reported with direct voice models that receive images themselves. By default, TellMeHow therefore hands the screen to a dedicated vision model and stores each of its descriptions. Every screen reading becomes verifiable and can serve as evidence.

**Can we stick to European providers?**

Partly, depending on the use. In TellMeHow, analysis, chat, and search can run on Mistral or on self-hosted OpenAI-compatible models, and screen text recognition for redaction runs on our servers. Real-time voice conversation currently relies on OpenAI, with Deepgram or ElevenLabs as options, providers that may process data in the United States: check their processing terms.

**How do we choose the models for an engagement?**

The TellMeHow Voice lab lets you talk with a configuration for three minutes, using a real engagement’s guide, and measures its latency and cost. Trials are kept and comparable, so you choose on numbers rather than on impressions. The chosen configuration then applies to the workspace, to a method, or to a single engagement.

---

- Français : https://tellmehow.io/fonctionnalites/ia-agnostique-et-labo-voix.md