1. Speech Recognition (TTS / ASR) Service

2. Speech Synthesis (STT) Services

What is STT and why is it needed?

Speech-to-Text (STT), also known as Automatic Speech Recognition (ASR), is the component that understands what the user is saying. It transcribes the caller's speech into text so that it can be processed by an LLM. Without accurate STT, the voice agent cannot know the user's request. For an in-depth comparison of STT and TTS providers, check our comprehensive STT/TTS selection guide.

Key considerations?

• Accuracy: Quality varies by model (general vs. phone-call optimized models, etc.). Good models achieve 5–10% WER, meaning the accuracy is 90–95%.
• Latency: For a live voice agent, low latency streaming transcription is needed so you're not keeping the user waiting. Represented by the Real-Time Factor (RTF) metric.
• Language support: Important if your service needs to handle multiple languages. Some STT engines specialize or have better accuracy in certain languages.
• Customization: Some providers allow custom vocabularies or even custom acoustic models (training on specific data) to improve accuracy on proper nouns or industry jargon.
• Features: Extras like speaker diarization, punctuation capitalization, profanity filtering, etc., can add value in voice agent use-cases.

Usual billing models:

• Per minute of input audio – typically billed per minute (or per second) of audio processed. Most cloud STT services charge based on the length of the audio input.
• Often pay-as-you-go, with automatic volume tier discounts. Some have free monthly minutes.
• Streaming vs batch: Some providers price streaming (real-time) and batch (asynchronous file transcription) similarly, others may differ slightly.

Usual price ranges:

~$0.012–$0.024 per minute for major cloud providers at low volumes (roughly 1–2.4 cents per minute). High-volume discounts can bring costs below $0.010 per minute. Committed enterprise contracts can lower it further.

Key providers

• Deepgram (Nova 3): Real-time STT engine optimized for low latency and high accuracy. Offers usage-based pricing and on-prem deployment.
• AssemblyAI (Universal-2, Slam 1): API-first STT provider with enhanced models for call transcription, speaker diarization, and summarization.
• OpenAI Whisper / gpt-4o-transcribe: Whisper is an open-source model known for strong multi-language transcription, especially in noisy or accented speech.
• Speechmatics (Ursa 2): Known for strong multilingual support and accent robustness. Offers full on-premise deployments and real-time transcription.
• Gladia (AI Solaria): STT startup focused on low-latency transcription with speaker detection and voice activity tracking.
• Google Cloud Speech-to-Text: Enterprise-grade STT with wide language support, streaming and batch modes, and enhanced phone call models.

3. Large Language Models (LLMs)

What is it and why is it needed?

The LLM is essentially the "brain" of the voice agent. It takes the transcribed text input and processes language – understanding intent and context – then generates a text response. In a modern AI voice agent, an LLM (like GPT-style models) enables natural conversation, dynamic responses, and handling of free-form user input that rule-based systems can't. For detailed guidance on selecting the right model, see our comprehensive LLM selection guide.

Key considerations:

• Model capability: Different LLMs have different strengths and weaknesses. Choosing an appropriate model impacts how well your agent can handle complex queries.
• Context length: Newer models in 2025 support very long inputs (sometimes millions of tokens), but using long contexts can be costly.
• Latency: Large models can be slow. For voice agents, you may favor a slightly smaller or distilled model if it significantly improves response time.
• Privacy & Hosting: Sending conversation data to an external API might raise compliance issues if the content is sensitive. For businesses handling sensitive data, review our guides on SOC 2 compliance and legal compliance requirements.
• Customization: Some providers allow fine-tuning (which may incur training and hosting costs).

Usual billing models:

Token-based billing– this is standard for most LLM APIs. You pay for input tokens (the text you send in, including conversation history) and output tokens (the text the model generates). A token is roughly 0.75 words, so 1,000 tokens ~ 750 words.

Key providers:

• OpenAI: Providers of GPT-4.1/GPT-4o/GPT-4o mini/GPT-4.1 mini, still a leader in capability and widely used.
• Google: Providers of Gemini 2.0 Flash, 2.0 Flash Lite, and 2.5 Flash. Known for ultra-fast, low-latency models optimized for enterprise use.
• Anthropic: Their Claude series models are known for handling long context and have a pricing model similar (per million tokens).
• Meta: Maintains the LLaMA series, currently at LLaMA 3.3. Meta's open-source approach powers many custom deployments.

4. Voice Agent Platform (e.g., VAPI, Bland AI)

What is it and why is it needed?

A Voice Agent Platform is an orchestration layer or framework for building the actual voice bot/agent. It typically ties together the telephony, STT, LLM, and TTS components, handling the call flow logic, state management, and integration with any backend systems.

Key considerations:

• Integration vs. all-in-one: Some platforms require you to bring your own STT, LLM, etc., while others provide an end-to-end solution.
• Features: Look for call recording, DTMF for IVRs, real-time handoff, latency management, and analytics.
• Scalability and Reliability: Handling potentially many concurrent calls is non-trivial. Platforms take care of scaling the voice infrastructure.
• Quality Assurance: Implementing proper testing and monitoring is essential for production deployments. Learn more about QA metrics and testing tools for voice agents.

Usual billing models:

Usage-based (per minute of call) – The platform might charge you per minute of voice call handled by the AI agent. This often encompasses the underlying costs (STT, TTS, etc.), essentially as a bundled rate.

Usual price ranges:

On the order of $0.05–$0.15 per minute for fully-managed voice AI platforms. Around 5–10 cents per minute is a typical ballpark for many platforms that bundle the costs.

Key providers:

• Vapi.ai: A developer-oriented platform (middleware style). You integrate your own models/services through it. Offers strong developer tools and flexibility.
• Bland AI: An end-to-end infrastructure-level platform. It provides its own integrated stack (STT, LLM, TTS all included and optimized together).
• Retell AI: Another voice AI platform in this space (similar segment as VAPI). Often mentioned alongside VAPI for bring-your-own-model setups.

5. Telephony / SIP

What is it and why is it needed?

This component handles the phone call itself – connecting calls over the Public Switched Telephone Network (PSTN) or via Session Initiation Protocol (SIP). Essentially, it's the layer that deals with phone numbers, call routing, and audio streaming between your user and the voice agent.

Usual billing models:

• Per-minute billing for calls. You're charged for the call duration. Rates typically differ for inbound vs outbound and by destination country/number type.
• Phone number rental fees: If you need dedicated numbers, there's a monthly fee per number. In the US this is around $1 per number per month.

Usual price ranges:

In the U.S., typical voice call costs are on the order of $0.005–$0.02 per minute. Toll-free numbers: receiving a call on a toll-free line might be around $0.02/min.

Key providers:

• Twilio: A market leader in cloud telephony APIs. Offers a wide range of voice features and global reach.
• Vonage: Another major CPaaS provider with similar global capabilities. Often slightly different pricing and has per-second billing by default.
• Others: Plivo, Telnyx, Bandwidth, SignalWire – these are other telephony API providers that sometimes offer lower pricing or specialized services.

Summary: AI Voice Agent Cost per Minute

Note: LLMs are priced per token (not per minute). This table provides estimated per-minute costs based on average token consumption in real-time voice interactions. Similarly, TTS services often charge per character, not per minute. These figures are meant for cost modeling, not direct billing rates.