> ## Documentation Index > Fetch the complete documentation index at: https://docs.lokutor.com/llms.txt > Use this file to discover all available pages before exploring further. # Lokutor > Voice agents, text-to-speech, speech-to-text and noise cancellation on Lokutor's own speech models, paid per use from a prepaid balance Lokutor builds its own speech models and serves them as a voice API: voice agents that hold a conversation over the web or the phone, text-to-speech, speech-to-text and noise cancellation. You pay per use from a prepaid balance in US dollars, with no plan and no monthly fee. A voice agent costs about 5¢ a minute, language model included. The exact prices are in [Pricing and billing](/pricing), and the live rate card is public JSON at [`GET /v1/pricing`](/api-reference/rate-card). ## What you can build 1. **Voice Agent**: full-duplex conversations over a single WebSocket, on the web or on a phone number. Speech recognition, the language model, the voice and turn-taking run together, with barge-in support. 2. **Streaming TTS**: send text over a WebSocket and receive audio as it is generated, with the first audio in a few hundred milliseconds. A one-shot `POST /tts/synthesize` returns a WAV file. 3. **Speech-to-Text**: transcribe a recorded clip or a live microphone stream. 4. **Noise removal**: clean a recording with `POST /denoise`. Voice agents clean the caller's audio on every call. ## What runs a call | Job | Model | Notes | | :- | :- | :- | | Speech synthesis | **Versa** (`versa-2.0`) | Lokutor's own. 10 voices, 9 languages. | | Turn-taking and barge-in | **Turno** | Lokutor's own. Decides from the audio whether the caller has finished or is interrupting. | | Noise suppression | **Psst** | Lokutor's own. Cleans the call before recognition. | | Speech recognition | NVIDIA's open **Parakeet** model | Runs on Lokutor's servers. Covers English, Spanish, Portuguese, French, Italian and German; not yet Catalan, Galician or Basque. | The models are small and efficient, and the hosted API serves them on GPUs. When call audio has to stay on your own infrastructure, the full models also run [on-premise](https://lokutor.com/on-premise-voice-ai/). Ito (text-to-speech) and Oído (speech recognition) are the open models for ESP32-class microcontrollers, the [on-device](https://lokutor.com/edge/) tier. See [Voices, Languages & Models](/voices-languages-models) for the voice list and the languages each part covers. ## Audio format | Property | TTS | Voice Agent | | :- | :- | :- | | **Input** | Text | Int16 PCM, 16 kHz, mono | | **Output** | Int16 PCM, 44.1 kHz, mono | Int16 PCM, 44.1 kHz, mono | | **Connection** | WebSocket (`/ws/tts`) or one HTTP request (`POST /tts/synthesize`, WAV) | Persistent WebSocket (`/ws/agent`) | The SDK handles resampling, buffering, and scheduling. ## Building with an AI assistant? These resources let a coding assistant work from the current documentation instead of from memory: | Resource | What it is | | :- | :- | | [API MCP server](/guides/mcp-server) | `https://api.lokutor.com/mcp`: voices, languages, models, limits and speech synthesis as MCP tools | | [Docs MCP server](/guides/mcp-server#lokutor-docs) | `https://docs.lokutor.com/mcp`: search and read these docs | | [Agent Skill](/guides/agent-skill) | `npx skills add https://docs.lokutor.com` installs a skill that teaches an agent how to integrate Lokutor | | [`/llms.txt`](https://docs.lokutor.com/llms.txt) | An index of every page with its description | | [`/llms-full.txt`](https://docs.lokutor.com/llms-full.txt) | Every page in one text file | | [OpenAPI spec](/api-reference/openapi-spec) | The HTTP API as OpenAPI 3.1 | ## Quick links Make your first API call in under 5 minutes. Create agents, manage API keys, and view usage. The prepaid balance, the rate card, and what a call costs. JS and Python SDKs. REST and WebSocket API docs. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.