Language meets domain
We focus on Korean STT adaptation, service terminology and patterns of speech. The company reports 16,000 hours of development audio; dataset size is distinct from recognition accuracy.
Conversational AI technology
1. Every word. Beyond the noise.
A conversation starts with listening. Short utterances, unfamiliar names and telephone audio all need to become text the next stage can use.
Input
Microphone or telephone audio
Output
Text, timestamps and speaker cues
Domain-adapted recognition, fine-tuned for Korean speech, noisy spaces, older voices and telephone audio.
A streaming conversation
Process audio in segments, distinguish partial from final transcripts, and tune voice activity detection and endpointing. The goal is to balance responsiveness with a conversation that does not cut people off.
Models that fit the context
An adapter layer can connect our adapted models or a customer-selected STT engine. Language, hosting requirements and latency determine the choice, verified on the same evaluation audio.
In practice
- “I’d like to change my booking for this Friday.”
- A transcript that preserves dates and intent
Service integration
Discuss your deploymentStart with real conditions
Separate 8 kHz telephone audio from higher-rate microphone input. Build test sets for noise, reverberation, speech rate and older speakers.
Connect recognition to work
Connect web, kiosk and call-centre audio to the conversation orchestrator, with transcripts, timestamps and error events. Integration scope is agreed for each deployment.
Evaluation
The same data, conditions and measurement boundaries. A result starts with a reproducible protocol.
Accuracy, by language
Report Korean CER alongside WER with fixed spacing, punctuation and number-normalization rules. Evaluate names and task-critical entities separately.
Latency, by stage
Evaluate first partial latency, finalization after speech ends and processing throughput at p50 and p95. State network boundaries and concurrency.
Streaming word error rate · WER ↓
UNVALIDATED TARGET
PROTOCOL
Re-run reference models on identical audio, transcripts and normalization. Published figure is not Korean-specific.
Korean character error rate · CER ↓
UNVALIDATED TARGET
PROTOCOL
At least 1,000 held-out utterances. Break out noise, older voices and 8 kHz audio; report 95% confidence intervals.
First transcript latency ↓
UNVALIDATED TARGET
PROTOCOL
Client audio send to first non-empty interim transcript. Report connection setup separately.
References
1. Deepgram Nova-3 · vendor benchmark2. OpenAI Whisper large-v3 · evaluation notesGoogle Cloud · evaluation guidance
Streaming word error rate · WER ↓ · Deepgram Nova-3 6.84%¹
Korean character error rate · CER ↓ · Whisper large-v3 uses CER for Korean²
First transcript latency ↓ · Matched-condition evaluation planned
2. Better answers begin with evidence.
Connect retrieval, language models and business systems. The aim is an answer grounded in approved information, with a clear boundary around what the assistant can do.
Input
Transcript, context and approved knowledge
Output
Grounded answers and approved actions
Connecting responses to relevant company documents and verified knowledge.
Documents into usable knowledge
Design chunking around document structure, metadata, permissions and versioning. Choose keyword retrieval, vector retrieval and reranking around the corpus and the questions.
The right model for the task
We distinguish model development from platform development. External LLMs are connected around task, speed and cost, while our platform coordinates retrieval, context and conversation history.
A policy for uncertainty
When evidence is insufficient, design for clarification, abstention or handover. Validate tool calls against allowed actions and add confirmation and audit records for consequential operations.
In practice
- Retrieve campus policy
- verify the source
- explain it in context
- connect to the right office
Service integration
Discuss your deploymentConnect knowledge and permissions
Connect approved FAQs, product information, academic rules and manuals. Define permissions, deletion and freshness policies before integrating retrieval and generation.
Fit the operating environment
Choose data location and model routing for SaaS, dedicated cloud or on-premise requirements. Define observability for conversation outcomes, cost and latency.
Evaluation
The same data, conditions and measurement boundaries. A result starts with a reproducible protocol.
Evaluate retrieval and generation separately
Measure evidence retrieval through Recall@k and ranking quality, then evaluate generated claims through faithfulness and citation precision.
Test when not to answer
Include unanswerable questions, changed policies and restricted documents. Combine automated evaluation with human-reviewed outcomes.
Retrieval recall · Recall@5 ↑
UNVALIDATED TARGET
PROTOCOL
≥300 human-labelled queries. Fraction of relevant evidence retrieved in the top five.
Claim groundedness ↑
UNVALIDATED TARGET
PROTOCOL
Share of verifiable claims supported by sources. Automated evaluation plus human review. Do not convert vendor 1–5 scales into percentages.
Citation precision ↑
UNVALIDATED TARGET
PROTOCOL
Check whether each cited passage actually supports the associated claim.
Retrieval latency ↓
UNVALIDATED TARGET
PROTOCOL
Target protocol: 100k fixed chunks, top-5, 10 concurrent requests. Generation latency is measured separately.
References
1. Microsoft · RAG evaluators2. Amazon Bedrock · evaluation metricsRAGAS · EACL 2024
Retrieval recall · Recall@5 ↑ · Azure: NDCG and retrieval evaluation¹
Claim groundedness ↑ · AWS: Faithfulness / Azure: Groundedness
Citation precision ↑ · AWS: Citation precision²
Retrieval latency ↓ · Matched-condition evaluation planned
3. Give a good answer a natural voice.
Design the moment an answer becomes sound. Naturalness, pronunciation, speaker consistency and time to first audio all shape the experience.
Input
Answer text, voice and speaking style
Output
Streaming speech
In-house speech synthesis and short-reference voice cloning, developed around natural conversation.
A voice at the core
Our own TTS development supports service-specific voices. The company-reported 3–7 second reference voice-cloning capability requires rights and consent; output quality must be evaluated across reference conditions.
From pronunciation to pacing
Normalize numbers, dates, abbreviations and borrowed words for speech. Shape sentence length and pauses, with pronunciation dictionaries and review sentences for important terminology.
Speech with less waiting
Design sentence-level synthesis before the entire answer is complete. Separate inference, encoding, transport and playback to optimize response time and support interruption.
In practice
- Deliver the same guidance with clear pronunciation and well-placed pauses.
Service integration
Discuss your deploymentMatch voice and channel
Validate codec and output levels for telephone bandwidth, kiosk speakers and mobile playback. Check voice consistency across channels.
Connect the appropriate engine
External TTS can complement language or hosting requirements. Design a shared interface for voice selection, style controls, stream completion and errors.
Evaluation
The same data, conditions and measurement boundaries. A result starts with a reproducible protocol.
Quality judged by listeners
Evaluate naturalness MOS, speaker similarity, pronunciation errors and preference separately, with randomized blind listening, native listeners and confidence intervals.
Separate synthesis and playback latency
Model latency and time to first audio are different measures. Record TTFB, real-time factor and playback readiness alongside codec, text length and concurrency.
Model inference latency ↓
UNVALIDATED TARGET
PROTOCOL
Development goal. Measure internal model time on short inputs; disclose hardware. Excludes network, app and playback buffering.
Client time to first audio byte · TTFB ↓
UNVALIDATED TARGET
PROTOCOL
Match region, language, codec, input length and concurrency. Also track a p95 goal of ≤500 ms.
Mean opinion score · MOS ↑
UNVALIDATED TARGET
PROTOCOL
Azure figure is guidance, not an Azure measured score. ≥100 sentences, 30 native listeners, randomized blind testing and confidence intervals.
References
1. ElevenLabs · latency definitions2. Microsoft Azure · TTS evaluation guidance
Model inference latency ↓ · ElevenLabs Flash v2.5 ~75 ms¹
Client time to first audio byte · TTFB ↓ · Different from the 75 ms inference figure
Mean opinion score · MOS ↑ · Azure recommended threshold >4.0/5²
4. Read the moment. Shape the voice.
Explore emotional cues, then adapt the response and its delivery. EMOTION is about finding an appropriate way to speak, without claiming to know a person’s inner state.
Input
Text and optional acoustic cues
Output
Context-sensitive response and delivery
Interpreting emotional cues in voice, facial expressions and text.
Read cues, preserve uncertainty
Design for cues such as excitement, worry, frustration or calm. Acoustic analysis requires a separately integrated and validated model. An emotion estimate is neither a diagnosis nor a fact about the person.
From acknowledgment to action
Define response policies: calm guidance for worry, concise next steps for frustration and appropriate warmth for excitement. Useful, accurate guidance comes before emotional performance.
Coordinate language and delivery
Control rate, pitch, pauses and intensity within the service design. Supported styles differ by engine and voice, so integration must distinguish external model features from our own development scope.
In practice
- “It’s my first time and I’m a little worried.”
- worry cue
- reassuring guidance
- measured pace
Service integration
Discuss your deploymentAdapt to the sentence and situation
Tune expression and speaking pace around the sentences and situations used in a service. Design delivery around the purpose of each response, from guidance to reassurance and confirmation.
Connect a production engine
An administrator can connect an emotion-response API for analysis, replies and synthesized audio. Validate contextual appropriateness and end-to-end latency by language and speaker.
Evaluation
The same data, conditions and measurement boundaries. A result starts with a reproducible protocol.
Does expression match intent?
Present the same text in multiple styles. Use blind listening to assess expression and appropriateness, recording class counts, rater counts, agreement and a confusion matrix.
Does speech quality hold up?
Check pronunciation and speaker consistency across styles, and watch for excessive expression. Evaluate listener preference and task success. No measured public score is claimed here.
5. A face for the conversation.
Connect speech to a digital persona that people can engage with on screen. Beyond appearance, the essentials are conversational timing and a useful service.
Input
Response audio and a persona image
Output
Synchronized avatar and service interface
Bringing voice and expression together in an approachable digital presence.
From image to presence
We develop photo-based avatars connected to speech, using images with appropriate rights. Design the persona, framing and interface together so people understand they are interacting with AI.
Speech and movement in time
Synchronize speech, pauses and mouth movement, with clear listening, processing and speaking states. Evaluate visual stability and consistency between frames.
Interaction for each channel
Design for viewing distance and input method across kiosks, web and exhibitions. Include a text-and-voice fallback when rendering is delayed.
In practice
- Listening
- preparing a response
- synchronized speech
- ready for the next question
Service integration
Discuss your deploymentDesign and engine, together
Agree on brand, persona rights, resolution and network conditions before choosing production and delivery methods. Identify concept imagery where no deployment photograph exists.
Connect to the conversation
Connect STT, RAG and TTS events to interface states. When a user starts speaking again, coordinate playback interruption and response updates.
Evaluation
The same data, conditions and measurement boundaries. A result starts with a reproducible protocol.
Visual quality and synchronization
Evaluate audiovisual offset, face and background consistency, and dropped frames separately, alongside human viewing tests. No quantified head-to-head result is claimed.
End-to-end conversation latency
Measure first rendering time and audiovisual synchronization. Fix resolution, frame rate, network conditions and concurrent sessions for reproducible evaluation.