Fixie AI

Processes raw audio input directly through an end-to-end trained speech-native model (ultravox-v0.7) that preserves tone, cadence, pitch, and emotional prosody without intermediate text transcription. Outputs audio responses with integrated text-to-speech, enabling natural conversational flow at sub-second latencies. The model operates on dedicated, purpose-built inference infrastructure managed by Ultravox, not via external LLM API calls.

Provides built-in, pre-configured integrations with 'largest telephony providers' (specific providers not named in documentation) to route inbound and outbound calls directly to Ultravox voice models. Handles SIP, PSTN, and VoIP protocols transparently; developers configure telephony routing via REST API without managing carrier connections or call signaling directly.

Exposes Ultravox voice models via REST APIs and native SDKs for web, mobile (iOS/Android), and backend platforms. Supports both request-response (single turn) and WebSocket streaming (continuous conversation) patterns. SDKs handle audio encoding/decoding, session management, and error handling transparently; developers interact with simple function calls rather than raw HTTP.

Ultravox v0.7 model runs on dedicated, purpose-built inference infrastructure optimized for sub-second response times. Published benchmarks show response latency faster than GPT-4, Gemini Live, and Claude Sonnet 4.5 on Big Bench Audio tasks (84% pass rate at fastest latency tier). Latency is a first-class optimization metric; specific millisecond latencies not published, but positioning emphasizes speed over accuracy trade-offs.

Ultravox uses a simple per-minute pricing model ($0.05/minute for all usage including TTS) with concurrency limits tied to subscription tier. Free tier: 5 concurrent calls; Pro tier: $100/month (annual) with higher concurrency; Enterprise: custom concurrency and pricing. Metering is transparent and usage-based — no per-call, per-token, or per-interaction surcharges documented.

Ultravox maintains conversation context across multiple turns within a session, enabling the model to reference prior messages and maintain coherent dialogue. Implementation details (context window size, session persistence, state management) are not documented. Appears to support continuous conversation without explicit context resets, but no information on how context is managed across calls or sessions.

Handles raw audio input (PCM, WAV, or streaming via WebSocket) and generates audio output via integrated text-to-speech (TTS) without requiring external TTS services. Audio encoding/decoding is abstracted by SDKs; developers work with audio streams or files without managing codec details. TTS is included in the per-minute pricing ($0.05/min), not a separate charge.

Ultravox v0.7 is benchmarked on Big Bench Audio, a standardized evaluation suite for speech AI models. Published results show 84% pass rate at fastest latency tier, positioning the model's accuracy/latency trade-off vs. competitors (GPT-4, Gemini Live, Claude Sonnet 4.5). Benchmarks are public and reproducible, enabling developers to evaluate suitability before committing.

Ultravox exposes call lifecycle events (call initiated, call ended, transcription available, etc.) via webhooks, enabling asynchronous integration with external systems. Developers configure webhook URLs in the API; Ultravox sends HTTP POST requests with call metadata and events. This enables decoupled workflows where voice interactions trigger downstream processes (CRM updates, logging, notifications) without blocking the call.

Ultravox can optionally extract and return conversation transcripts (text representation of audio dialogue) via API responses or webhooks. Transcripts are generated from the speech-native model's internal representation (not via separate speech-to-text); transcript availability and format are not fully documented. Transcripts enable logging, compliance, and debugging without requiring separate transcription services.