Integrating Asterisk/FreePBX with AI for Call Handling

How to Integrate Asterisk with AI?

"We have Asterisk 18, want to add a voice assistant, but ready-made solutions bring a cloud subscription and don't let us control latency." This is a typical pain point. The default FreePBX capabilities — DTMF menus and playing pre-recorded files. AI enables a dialogue: user speaks, system understands, LLM makes a decision, TTS responds. But integration is not "install a module". It's about choosing the stack (STT: Whisper or Deepgram? LLM: LLaMA 3.1 70B or GPT-4o-mini? TTS: Piper or ElevenLabs?), about latency p99 and dialog convergence. Here's how we build such projects on-premise or hybrid.

What Problems We Solve

Problem 1: Recognition quality in noisy environments. Call centers, open-plan offices, street noise — standard STT models give WER > 30%. We use Whisper large-v3 with fine-tuning on your audio corpus (300 hours of recordings reduce WER to 10%) or Deepgram Nova-2 with noise suppression on input.

Problem 2: End-to-end latency > 2 seconds. Users won't wait longer than 1.5 s. We optimize the pipeline: streaming STT with partial results, early-exit LLM, TTS generating first packet in 150 ms. On the Whisper + vLLM + Piper stack we get p99 = 1.1 s.

Problem 3: Fault tolerance. An AI service crash should not break the call. We design a fallback: on AI node timeout, the dialplan switches to standard IVR. The ARI application monitors each component's health and on error returns control to Asterisk.

How We Do It: Stack and Configs

The choice of interface — AGI or ARI — depends on real-time requirements and dialog complexity. Comparison below.

Example AGI script (Python):

# agi_bot.py
import sys
from asterisk.agi import AGI

agi = AGI()
agi.answer()
agi.set_variable("CHANNEL(audioreadformat)", "slin16")

# Record audio from user
agi.record_file(
    "/tmp/user_audio",
    "wav",
    "#",        # stop key
    3000,       # timeout ms
    0,          # offset
    True,       # beep
    3           # silence threshold
)

# STT + LLM + TTS in separate service
import requests
with open("/tmp/user_audio.wav", "rb") as f:
    stt_response = requests.post("http://ai-service/stt", files={"audio": f})
transcript = stt_response.json()["text"]

llm_response = requests.post("http://ai-service/chat",
                              json={"text": transcript})
response_text = llm_response.json()["response"]

tts_response = requests.post("http://ai-service/tts",
                              json={"text": response_text})
with open("/tmp/response.wav", "wb") as f:
    f.write(tts_response.content)

agi.stream_file("/tmp/response")

ARI example (asyncio):

import asyncio
import aiohttp
from ari_client import ARIClient

async def handle_stasis(channel_id: str, ari: ARIClient):
    """Incoming call handler via ARI"""
    await ari.answer(channel_id)

    # Create audio snapshot of the channel
    await ari.channel.record(
        channelId=channel_id,
        name=f"call_{channel_id}",
        format="wav",
        terminateOn="silence",
        maxSilenceSeconds=2
    )

Why Choose AGI Over ARI?

If your scenario is simple voice I/O (e.g., "say your name, the system will find the client"), AGI gives minimal development time. Writing the script takes 1-2 days, debugging another week. ARI requires async architecture and session management but allows handling barge-ins and partial recognition results. For contact centers with natural dialog, ARI is the only choice.

How to Achieve Latency < 1 Second?

We build the pipeline:

1. STT — Whisper medium.en + ModelScope (INT8 quant) on GPU T4 — first token in 250 ms.
2. LLM — vLLM with LLaMA 3.1 8B, prefill at 20 tokens, early stop generation — 400 ms. Compare: vLLM is 2.5x faster than standard Hugging Face pipeline under load.
3. TTS — Piper (VITS) with first packet synthesis in 100 ms. Result: 750 ms end-to-end. Under load of 10 simultaneous calls, latency p99 stays at 1.2 s.

According to Asterisk documentation, for streaming audio it is recommended to use ARI, as AGI does not support full-duplex transmission.

LLM Comparison for Voice Scenarios

name: "llama_ensemble"
backend: "ensemble"
input [
  {
    name: "text_input"
    data_type: TYPE_STRING
    dims: [ -1 ]
  }
]
output [
  {
    name: "text_output"
    data_type: TYPE_STRING
    dims: [ -1 ]
  }
]

Process

1. Analysis (1–3 days): Audit current Asterisk/FreePBX config, gather scenario requirements, measure average call duration.
2. Design (2–5 days): Choose stack (models, frameworks), design integration, prototype dialog.
3. PoC (1–2 weeks): Deploy on test PBX with 1 inbound number. Demo to client.
4. Production (2–4 weeks): Deploy on target servers, set up monitoring (Prometheus + Grafana), integrate with CRM.
5. Support (2 weeks after launch): Train operators, tweak dialog script, optimize latency.

Timelines and Cost

• Basic AGI integration (one scenario, Whisper + LLaMA + Piper): from 2 to 3 weeks.
• Full ARI implementation (real-time dialog, multilingual, fault tolerance): from 1 to 1.5 months. Cost is calculated individually after infrastructure audit. Get a consultation — we estimate your project in 1 day.

What's Included

• Deployment and operation documentation.
• Repository with configs and scripts (Git).
• Operator training (2 sessions of 1 hour).
• Two weeks of post-launch support.
• 3-month warranty on code.

We are a team with 5 years of experience in AI integrations for telephone systems. We have 12+ projects for contact centers up to 50 lines. Certified Asterisk and FreePBX engineers (Sangoma distribution). Contact us — we will send a technical and commercial proposal.

Learn more about Asterisk

Speech Recognition and Synthesis: ASR, TTS, Voice Cloning

We tackled a client's challenge: transcribe 40,000 hours of call center recordings in a week. Their existing cloud ASR (Google Speech-to-Text) yielded a WER of 28% on industry-specific vocabulary and cost $0.006 per minute — prohibitively expensive at that volume. The goal was to reduce WER below 10% and switch to self-hosted inference. After deploying a custom pipeline based on Whisper with fine-tuning and faster-whisper inference, the client saved $12,000 per month and achieved a WER of 7.3%.

How does speech recognition ASR handle noisy call center recordings?

The most common issue is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec. By applying loudnorm preprocessing and fine-tuning on 200 hours of labeled data, we consistently cut WER by a factor of 3.

Typical problems we encounter

WER does not converge to the desired metric. Often the culprit is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec.

Diarization fails with more than two speakers. pyannote/speaker-diarization-3.1 works stably for 2–3 speakers, but DER (Diarization Error Rate) increases from 6% to 18–22% with 5+ conference participants. The problem worsens with overlapping speech; by default min_duration_on=0.1 cuts short interjections. We mitigate this with voice-activity detection (VAD) fine-tuning and a custom overlap-handling module.

Voice cloning — latency vs. quality. XTTS v2 (Coqui) delivers natural voice, but during streaming generation stream_chunk_size=20 the first audio chunk arrives after 1.4–2.0 seconds — unacceptable for interactive scenarios. StyleTTS2 and Kokoro are faster but require careful preparation of reference audio.

How do we solve it in practice?

The basic stack for a production pipeline:

• ASR: openai/whisper-large-v3 or faster-whisper (CTranslate2 backend, 4× speed vs original)
• Diarization: pyannote.audio 3.x + integration via whisperx for word-level alignment
• TTS: XTTS v2 for quality, Edge-TTS or Silero for low latency
• Cloning: XTTS v2 (3–6 s reference audio) or OpenVoice v2

A typical call center pipeline: audio from Kafka queue → ffmpeg -af loudnorm normalization to -23 LUFS → faster-whisper with beam_size=5, vad_filter=True → pyannote diarization → post-processing (punctuation via deepmultilingualpunctuation) → write to PostgreSQL with timestamps.

Case study from our practice. A fintech company with 12,000 calls per day. Initial WER on Russian with banking vocabulary — 22% (Google STT). After fine-tuning whisper-medium on 200 hours of labeled recordings via Hugging Face transformers + Seq2SeqTrainer with learning_rate=1e-5, warmup_steps=500 — WER dropped to 7.3%. Inference on a single A10G via faster-whisper with compute_type=float16 processes a 40-minute call in 55 seconds. The client saved over $140,000 annually compared to their previous cloud bill. Contact us for a free pilot estimate to see similar savings on your data.

How to fine-tune Whisper on domain data?

When a general model underperforms, fine-tuning is the first tool. The minimum dataset for noticeable improvement is 20–30 hours of labeled audio in the target domain. Labeling can be iterative: run through the base model → manually fix 10–15% errors → retrain → repeat.

training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=16,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    fp16=True,
    predict_with_generate=True,
    generation_max_length=225,
)

Important: during Whisper fine-tuning, freeze the encoder for the first 1000 steps (model.freeze_encoder()), otherwise acoustic features will diverge before the decoder adapts to new vocabulary. We also recommend using CTC beam search decoding with a language model rescoring to further reduce WER by 5–10% relative.

Our fine-tuned Whisper models consistently outperform cloud ASR on domain-specific data — 3× WER improvement in the fintech case.

Speech synthesis: How to choose a model for your task?

For interactive bots requiring TTFB < 300 ms — Silero or Kokoro. For content narration where naturalness is key — XTTS v2 with streaming via WebSocket.

Our process and deliverables

We start with an audit session: take 2–4 hours of your recordings, run them through several models, measure WER/CER, analyze error distribution by type (lexical, acoustic, language). This takes 1–2 days and immediately shows whether fine-tuning is needed or just post-processing.

Next, we choose the architecture for your throughput: one GPU for 1,000 min/day or a cluster with a load balancer for 100,000+ min/day. Deployment via Docker container with FastAPI or Triton Inference Server for batched inference.

What you get after engagement:

• Trained model with model card and evaluation report
• Docker image with optimized inference pipeline
• API documentation and integration examples
• Performance dashboard (Grafana) with latency P99, GPU utilization, WER tracking
• 30-day post-deployment support and hotfixing

Timelines depend on complexity:

• Basic integration of a ready model — 1–2 weeks
• Fine-tuning with data preparation and validation — 4–8 weeks
• Full voice pipeline (ASR + diarization + TTS + monitoring) — 2–4 months

Project investments typically range from $20,000 to $80,000. Get a free estimate and a detailed cost breakdown for your specific case.

Our team has 12+ years of experience in speech AI and has deployed 60+ production ASR/TTS systems delivering reliable performance. Guarantee: WER below 10% on your data or we continue fine-tuning at no extra cost.

Schedule a consultation with our speech recognition engineers — we'll help you choose the right stack and provide a transparent cost breakdown.