Urdu Voice AI for Pakistani Call Centres and Field Teams
Pak-LLM's Urdu Voice AI converts speech to speech in real time: Faster-Whisper transcribes Urdu, Roman Urdu, and code-mixed speech, the language model answers, and neural TTS speaks the reply. Audio is processed ephemerally and is not retained after the session ends.
- Written for:
- Call centres, telcos, banks, e-commerce support desks, and field sales teams operating in Urdu or Roman Urdu.
- Last reviewed:
- Author:
- Saad Bin Farrukh
What does "real-time Urdu voice AI" actually mean in practice?
It means a caller speaks Urdu and hears an Urdu reply without a text step in between, with the system able to be interrupted mid-sentence. Pak-LLM chains streaming ASR, the language model, and streaming TTS so the first audio response begins before the caller finishes a long question.
- Handles Urdu script, Roman Urdu, and Urdu-English code-mixing in the same utterance.
- Supports barge-in: the caller can interrupt the agent and the response is truncated.
- Keeps a per-session transcript for the operator dashboard, not a permanent audio archive.
- Exposes the same capability through the web workspace and a REST/WebSocket API.
How is caller audio handled and retained?
Audio is transcribed in volatile memory and discarded when the session closes. Pak-LLM does not use customer call audio to train foundation models, and the zero-retention setting is the default rather than an enterprise add-on.
- No audio recordings are written to long-term storage under the default configuration.
- Text transcripts can be disabled entirely for regulated deployments.
- On-premise appliance deployments keep all audio inside the customer's network.
Which Pakistani languages and accents are covered?
The voice stack is tuned primarily for Urdu and Roman Urdu, with regional language coverage for Punjabi, Sindhi, Pashto, Balochi, Saraiki, and Kashmiri at varying maturity. Urdu and Roman Urdu are the production-grade pair; regional languages should be piloted before rollout.
- Urdu (Nastaliq) and Roman Urdu: production use.
- Punjabi Shahmukhi, Sindhi, Pashto, Balochi, Saraiki, Kashmiri: available, accuracy varies by domain.
- Arabic and English are supported for bilingual Gulf-facing operations.
How does pricing work for a voice deployment?
Voice usage is billed in Pakistani rupees against a token allowance rather than in USD per minute. The entry tier is PKR 950 per month with 5,000,000 tokens included; the flagship tier is PKR 6,500 per month with 50,000,000 tokens included.
- Genesis (v1.0): PKR 950/month, 5M tokens, PKR 150 per additional 1M tokens.
- Azaadi (v2.0): PKR 2,850/month, 18M tokens, PKR 300 per additional 1M tokens.
- Ummah (v3.0): PKR 6,500/month, 50M tokens, PKR 450 per additional 1M tokens.
- Enterprise on-premise licensing is quoted separately per appliance.
What does a realistic rollout look like?
A typical rollout starts with a single queue and a scripted knowledge base, runs in shadow mode beside human agents for two weeks, then moves to assisted mode where the agent drafts and a human approves. Full automation is only enabled for intents that clear an agreed accuracy threshold.
- Week 1–2: knowledge base ingestion and intent inventory.
- Week 3–4: shadow mode with human review of every transcript.
- Week 5+: assisted mode, then selective automation with escalation rules.
- Ongoing: weekly transcript review and prompt/guardrail updates.
Frequently asked questions
Does Pak-LLM store my call recordings?
No. Under the default configuration, audio is transcribed in memory and discarded at the end of the session. Transcript storage can be disabled for regulated deployments, and on-premise appliances never send audio outside your network.
Can the voice agent speak Roman Urdu instead of Urdu script?
Yes. The ASR and TTS layers handle Roman Urdu and Urdu-English code-mixing, and you can pin the reply script per deployment so agents consistently answer in the script your customers use.
What latency should I expect?
Pak-LLM targets sub-300ms time-to-first-token for text on its domestic routing tier. End-to-end voice latency also includes ASR and TTS, so measure speech-to-speech latency on your own network rather than relying on vendor figures.
Is Urdu Voice AI available on-premise?
Yes. The on-premise appliance bundles ASR, the language model, and TTS so no audio leaves the customer network. This is the option used by banking and defence-adjacent deployments.