Cloud AI services (Plugin)

Plugin (ai-cloud / ai-voicemail): Optional plugins from the official catalog. Enable them under Integrations → Plugins (see Plugin system & developer guide).

AI → Cloud services (admin only) holds the accounts of cloud AI providers in one place and chooses, for each job and language, which engine AI features use: a local model of this server (see Local AI models) or a cloud provider.

A cloud engine sends audio or text to that provider. Check its terms and your data-protection duties (KVKK/GDPR) before choosing one. Local models keep everything on the server.

Providers

Each card has the key fields, a link to get a key, a Test button (a cheap authenticated call that shows the provider's own error when the key is wrong) and this month's use: minutes of audio turned into text, characters spoken and language-model tokens (table ai_cloud_usage).

Keys are stored encrypted in sys_settings as ai_tts.<provider>.<field> — the names the Cloud TTS page used before, so existing accounts keep working. The Cloud TTS page now only shows the status of its providers and links here; changing keys needs the admin role.

Engine for each job

For speech, speech to text and the language model in Turkish, German and English, choose a running or downloaded local model or a configured cloud provider (ai_engine.<job>.<lang> in sys_settings, values local:<model id> or cloud:<provider>). AI features that come later (AI applications, voice requests) use these choices, so switching between local and cloud needs no other change.

Trying

AI → Local models → Try: speech to text lists the configured cloud providers next to the local models: record from the microphone or upload a WAV file, pick the language, and compare the text and the time each engine needs.

Language models

Large language models are not run on the PBX itself (most servers cannot carry them); they come from the providers above. They are used by:

• AI applications — routing by speech and voice requests with Understanding: cloud model (AI applications),
• voicemail as text — the one-sentence summary (below).

The model of Google AI, OpenAI, Groq and OpenRouter is fixed (a small, fast one: gemini-2.5-flash, gpt-4.1-mini, llama-3.3-70b-versatile, google/gemini-2.5-flash). For the three providers below the card has a Model field; Test sends one tiny chat request, so a wrong model name shows at once.

Voicemail as text

AI → Cloud services → Voicemail as text: a new voicemail is written down with the speech-to-text engine of the PBX's prompt language (a small local model such as Vosk or Whisper, or a cloud provider) and, if summary is ticked, summarised in one sentence by the language model of that language ("John Smith wants to move tomorrow's 10:00 meeting to Friday").

• The voicemail e-mail gets both under the list of details: template variable {ai_text} of the voicemail template (E-mail → Templates; a changed template needs {ai_text} added).
• The app notification shows the summary (or the start of the transcript) under the caller (Android 1.0.66).
• My Phone → Voicemail and the mobile apps' voicemail list show it next to the message.

Each message is processed once (table voicemail_ai); the e-mail comes first and the push uses the result. Voicemail is handled by Asterisk's own user, which cannot read the AI keys, so the work is done by the portal on the same server (api/voicemail_ai.php, only from 127.0.0.1, signed). Nothing here can hold back the e-mail or the notification: when the engine fails, they go out without the text. With a cloud engine the recording (speech to text) or its text (summary) is sent to that provider.