The Voice
Intelligence
Blog
Everything Voice AI: from the physics of audio codecs to the architecture of real time streaming pipelines. Written by builders, for builders.
Resolving Time Expressions: The Hidden Difficulty in Voice Scheduling
Booking by phone sounds simple until you convert what people say into a timestamp. Relative days, ambiguous words such as parso, half-hour forms such as dhaai, and festivals that move each year must all resolve to one exact slot.
Inverse Text Normalisation: Turning Spoken Words Into Written Records
Speech recognisers output words. Business systems need phone numbers, amounts, dates and email addresses in written form. Inverse text normalisation is the layer in between, and for Indian speech it is unusually difficult.
Prosody and Emphasis: Why Synthetic Speech Still Sounds Flat
Synthetic voices have largely solved pronunciation, yet many still sound flat. The reason is prosody: the pitch, timing and emphasis that carry meaning text does not encode. This is why it is hard and how researchers approach it.
Speaker Diarization: Separating Voices in Recorded Conversations
Diarization answers the question of who spoke when. It underpins call records, quality review and analytics, and telephone audio makes it unusually difficult through narrow bandwidth, codec artefacts, short turns and overlapping speech.
Acoustic Echo Cancellation: Why Voice Agents Hear Themselves
A voice agent speaks and listens on the same call at the same time, so its own voice returns on the incoming path. Without echo cancellation the agent transcribes itself, interrupts itself and can enter a feedback loop.
Emotion Recognition in Speech: Capabilities and Limitations
Speech emotion recognition promises to tell whether a caller is angry, anxious or satisfied. The reality is more modest and more interesting. Here is what the technology can do, where it fails, and how to use it responsibly.
Personal Data in Call Transcripts: Redaction, Consent and India's DPDP Act
Every customer call can capture names, phone numbers, addresses and identity numbers. With India's Digital Personal Data Protection Act now moving into force, businesses need a clear approach to consent, redaction and retention.
Retrieval-Augmented Generation Under Real-Time Constraints
Retrieval-augmented generation is the standard way to ground language models in business knowledge. On a live call, every retrieval decision competes with the silence a caller can hear. Here is how the design changes.
Grounding and Verification: Reducing Hallucinations in Voice Agents
A language model that invents a refund policy in a chat window is a problem. The same mistake spoken aloud on a customer call is worse. Here are the principles for grounding and verifying what a voice agent says.
Code-Switching and Speech Recognition: The Challenge of Hinglish
Most callers in urban India do not speak pure Hindi or pure English. They mix both, often inside a single sentence. Here is why that breaks conventional speech recognition, and how the field is approaching the problem.
Evaluating Conversational Voice Agents: Methods and Metrics
A voice agent can pass every unit test and still frustrate real callers. This guide covers component and end-to-end metrics, latency percentiles, simulated callers, LLM judges, audio testing and production experiments.
Text Normalisation for Indian Speech Synthesis: Numbers, Currency and Dates
A speech synthesiser cannot read ₹2,50,000, 05/06/2026 or GST without first turning them into words. Text normalisation for Indian languages brings lakhs, irregular Hindi numerals, code-mixing and more.
Backchannels and Barge-In: Managing Interruptions in Voice Conversations
Listeners constantly say hmm, haan and achha while someone else talks. A voice agent must separate these supportive signals from genuine interruptions, and recover gracefully when a caller really does cut in.
Prompt Injection Through Speech: Securing Voice AI Agents
A voice agent takes instructions in natural language from anyone who dials the number. That makes prompt injection a practical risk, not a theoretical one. This post explains how speech changes the threat and which defences actually hold.
Schwa Deletion and Indian Names: The Pronunciation Problem in Speech Synthesis
Hindi spelling hides a vowel that is often not pronounced. For speech synthesis this inherent schwa, along with retroflex consonants, aspiration and romanized names, turns pronunciation into one of the hardest problems in Indian voice AI.
Writing for the Ear: Prompt Design for Spoken Language Output
Language models learned to write from documents, so their default output is built for the eye: headings, lists, long clauses and parentheticals. A phone caller hears none of that structure. Here is how to prompt for speech instead.
Beyond Word Error Rate: Measuring Speech Recognition That Works
Word error rate is the standard metric for speech recognition, yet it can mislead badly for Hindi, Gujarati and code-mixed telephone audio. This post covers normalization, CER, entity accuracy, task success and building test sets you can trust.
Token Fertility in Indic Languages: Why Hindi Requires More Tokens Than English
The same sentence can consume far more tokens in Hindi or Gujarati than in English. This post explains subword tokenization, byte-level fallback and UTF-8, and why that imbalance affects context, speed and quality of language models.
The 200-Millisecond Gap: Turn-Taking in Human and Machine Conversation
Humans take turns in conversation with gaps of roughly a fifth of a second, far faster than they can plan speech from scratch. This post explains how that is possible and why voice agents must learn to predict, not just react.
Voice AIVoice AI for Bharat: Why Multilingual Agents Are the Next Big Leap
The next 500 million AI users will not primarily speak English. They will speak Hindi, Tamil, Marathi, Bengali, Telugu, and dozens of other languages. Building Voice AI for this audience requires rethinking assumptions made by Western AI research.
Voice SynthesisText to Speech in 2026: The Race for Zero Latency, Human Quality Voices
Modern neural TTS can clone a voice from 10 seconds of audio and synthesize speech faster than real time. Understanding how TTS systems work, and where they still fall short, is essential for anyone building Voice AI products.
AI and LLMsThe LLM Phone Call: How Language Models Are Rewriting Conversations
Putting an LLM on a phone call is harder than it looks. Context windows, conversation memory, function calling for real business actions, and tight latency constraints all collide. Here is how it is done right.
TechnicalCodecs Demystified: Why G.711 Still Rules the Phone World
Audio codecs are the hidden variable in every Voice AI deployment. They determine fidelity, bandwidth, latency, and ASR accuracy. Whether you are debugging a muffled AI call or optimizing for mobile networks, codec knowledge is essential.
ArchitectureReal Time Streaming Audio: How AI Voice Agents Speak Without Lag
The magic of a natural AI voice conversation comes down to latency. Understanding how streaming audio pipelines work, from audio capture to TTS playback, reveals the engineering beneath what feels like a simple phone call.
TelephonySIP, PSTN and the Phone Stack: What AI Voice Agents Actually Dial Into
Before an AI agent says a single word, its audio has to travel through a web of protocols and switches that date back to the 1950s. Understanding the PSTN, SIP, and the full telephony stack is essential for anyone building reliable Voice AI at scale.
Deep DiveVoice Activity Detection: The Silent Engine Behind Every AI Call
Voice Activity Detection is the invisible gatekeeper of every AI voice conversation. Without it, AI agents would talk over you, process silence as speech, and rack up unnecessary compute costs. Here is how it actually works.
Ready to put Voice AI to work?
Deploy your AI voice agent in under a minute. No developers required.