Introduction

With Asterisk 21, real-time speech-to-text (STT) is more accessible than ever, allowing seamless call transcriptions, IVR automation, and AI-driven call routing. I’ve tested three STT solutions—AEAP, Google Speech-to-Text, and Vosk—each with unique strengths and trade-offs.

In this guide, I’ll compare installation, latency, cost, and accuracy to help you choose the best option.

---

1. AEAP (Asterisk External Application Protocol) Speech-to-Text

AEAP is Asterisk’s built-in module that enables speech recognition by connecting to external services.

Installation

AEAP is already included in Asterisk 21, but you need to enable it:

cd /usr/src/asterisk-21
make menuselect

• Navigate to “Resource Modules”
• Enable res_aeap.so
• Recompile and restart Asterisk

make && make install && systemctl restart asterisk

Configuring AEAP in Dialplan

Add this to your extensions.conf:

exten => 100,1,Answer()
 same => n,SpeechCreate()
 same => n,SpeechStart()
 same => n,SpeechBackground(beep,10)
 same => n,Set(SPEECH_TEXT_RESULT=${SPEECH_TEXT(0)})
 same => n,Verbose(1, "Recognized: ${SPEECH_TEXT_RESULT}")
 same => n,Hangup()

Pros & Cons

• No external API costs
• Native support in Asterisk 21
• Limited accuracy compared to AI-powered solutions
• No built-in support for custom language models

---

2. Google Speech-to-Text with Asterisk

Google’s Speech-to-Text API offers high accuracy and multilingual support but comes with usage-based pricing.

Installation

1. Create a Google Cloud project and enable Speech-to-Text API
2. Generate a Service Account JSON Key
3. Install Google’s SDK:

pip install google-cloud-speech

Configuring Google STT with Asterisk

Modify your extensions.conf to send audio to Google via an AGI script:

exten => 200,1,Answer()
 same => n,Record(/tmp/audio.wav,5,30)
 same => n,AGI(google-stt.agi,/tmp/audio.wav)
 same => n,Verbose(1, "Google Response: ${STT_RESULT}")
 same => n,Hangup()

google-stt.agi (Python Script for Google STT)

from google.cloud import speech
import sys

client = speech.SpeechClient()
audio = speech.RecognitionAudio(uri=sys.argv[1])
config = speech.RecognitionConfig(encoding="LINEAR16", language_code="en-US")

response = client.recognize(config=config, audio=audio)
print(f"STT_RESULT={response.results[0].alternatives[0].transcript}")

Pros & Cons

• Highly accurate
• Supports multiple languages
• Speaker diarization & noise filtering
• Expensive for high call volumes
• Requires stable internet connection

---

3. Vosk (Offline Speech-to-Text for Asterisk)

Vosk is an open-source, offline STT engine, making it a great option for self-hosted and low-latency applications.

Installation

git clone https://github.com/alphacep/vosk-api.git
cd vosk-api/python
pip install -r requirements.txt

Configuring Asterisk with Vosk

Use an AGI script to send audio to Vosk:

exten => 300,1,Answer()
 same => n,Record(/tmp/vosk_audio.wav,5,30)
 same => n,AGI(vosk-stt.agi,/tmp/vosk_audio.wav)
 same => n,Verbose(1, "Vosk Response: ${STT_RESULT}")
 same => n,Hangup()

vosk-stt.agi (Python Script for Vosk STT)

from vosk import Model, KaldiRecognizer
import wave
import sys

model = Model("model-en")
rec = KaldiRecognizer(model, 16000)

with wave.open(sys.argv[1], "rb") as wf:
    rec.AcceptWaveform(wf.readframes(wf.getnframes()))

print(f"STT_RESULT={rec.Result()}")

Pros & Cons

• Runs offline (no API costs)
• Lower latency than Google
• Customizable models
• Less accurate than Google STT
• Limited language support

---

4. Comparing Performance: AEAP vs Google vs Vosk

Feature	AEAP	Google STT	Vosk
Cost	Free	Paid (per min)	Free
Latency	Low	Medium	Low
Accuracy	Basic	High	Medium
Language Support	English only	120+ languages	Limited
Internet Required	No	Yes	No

---

Final Thoughts: Which One to Choose?

• Use AEAP if you want free, built-in recognition and don’t need high accuracy.
• Use Google STT if you need enterprise-level accuracy & multi-language support.
• Use Vosk if you want offline speech recognition without API costs.

For my projects, I’ve found that Google STT works best for customer service calls, while Vosk is great for VoIP-based AI assistants where cost matters.

Which STT engine are you using with Asterisk? Let me know in the comments!