Platform Overview
TopCalls connects phone lines (SIP/PSTN) to AI agents. We run the telephony, the AI pipeline and the audio. You decide what your agents say and do.What TopCalls Does
Telephony Infrastructure
SIP trunking, carrier registration, number provisioning and audio streaming.
AI Orchestration
Speech recognition, the conversation model and voice synthesis, tuned for phone calls.
Campaign Management
Outbound queues with retries and dispatch hours, plus inbound answering.
Call Analysis
Call summaries and structured data extraction with the fields you define.
System Architecture
Conversation Modes
Every call runs in one of two modes:live or legacy. New campaigns and API calls default to Live.
Live (default)
One voice model listens and speaks, with no separate speech recognition or voice synthesis step.
Live plays
background_audio at background_audio_gain, like Legacy.
Legacy
Separate speech recognition, language model and voice synthesis, each one configurable.Send
mode: "legacy". A request that leaves mode out runs Legacy when it sets a voice, a model, or anything from the Live “Not available” list.The Call Lifecycle
1
1. Call Creation
You place a call with
POST /v1/calls, or a campaign dispatches one. The request is validated and credit is held for the full max_duration at the per-minute rate.2
2. Dispatch
The call goes out to the phone network. Status moves from
queued to in_progress.3
3. Connection
The recipient picks up. Set
first_sentence to control the opening line: Legacy speaks it word for word, Live may reword it slightly. Without it, the agent decides how to open.4
4. Conversation
In Live, one voice model listens and speaks. In Legacy, speech recognition transcribes the caller, the language model writes the reply and voice synthesis speaks it. Either way the agent uses your knowledge base, calls your tools and can end the call when it’s done.
5
5. Completion
The call ends, from either side. TopCalls saves the transcript, fetches the recording (about 15 seconds later), writes the summary, extracts your
analysis_schema fields and maps them to an outcome with outcome_mapping.6
6. Webhook Delivery
Your server gets a webhook with the transcript, recording URL, summary, analysis data and your metadata. Values you sent under platform keys are ignored:
lead_context, max_duration and attempt_number always carry the platform’s values.Key Features
Voicemail Detection
When a call reaches voicemail, the agent hangs up andanswered_by reads voicemail. A call where the person spoke reads human, anything else unknown.
Function Calling
Give agents tools that reach your systems mid-call: book appointments, look up orders, update a CRM.Knowledge Base
Attach text entries with product, pricing and policy details. The agent draws on them during the call.Structured Data Extraction
Define a schema and get answers back from each call: “Did the customer agree to a demo?”, “What objections came up?”Multi-Language Support
Up to 78 languages with dialect control in Legacy. The exact set depends on the speech recognition and voice models you pick; the default voice model covers 44.Who Does What
You’re responsible for following local calling laws (TCPA/TSR/DNC, GDPR), including consent and opt-outs.
Next Steps
AI & Voice Customization
Instructions, model, voice and language.
Campaign Management
Outbound queues and inbound answering.