Speech Products: NLU, Neuron & Diarization
Transcription answers what was said. The products on this page answer the harder questions that follow—what it meant, who said it, and what to do about it. Natural Language Understanding (NLU) and Speaker Diarization convert raw conversation into structured signals you can route, score, and report on at scale: intent, summaries, topics, call outcomes, sentiment, language, and speaker attribution.
Each capability is exposed as a service you call on demand, so you can compose only the analysis a given workflow needs. NLU operates on text—typically the output of ASR & Transcription—while Speaker Diarization operates directly on audio.
Natural Language Understanding (NLU)
NLU applies natural language processing to supplied text to infer a speaker’s intent and extract meaning. It groups several subproducts, each addressing a distinct post-transcription need. All are requested through the NLU service.
Call Summarization
Turn full call transcripts into concise, consistent records—reducing after-call work and improving the quality of what lands in your CRM or case notes. The transcribed call text can drive any of the following three features, each requested independently:
Call Summary
Condense a conversation to its essentials. The output length is configurable, either by target word count or by number of bullet points.
Topic Detection
Identify and categorize the main subjects discussed—from billing questions to technical support and beyond—so calls can be classified and routed without manual tagging.
Outcome Detection
Record the result of each call—whether a sale closed, a task scheduled, or an issue left unresolved—making conversion and resolution measurable across large call volumes.
Sentiment Analysis
Sentiment Analysis uses the transcribed text to score the emotional tone of a single sentence or an entire call, surfacing dissatisfaction and escalation risk that a raw transcript hides. It runs as a batch process and returns one of seven graded values:
- Positive
- Mostly Positive
- Mixed
- Neutral
- Mostly Negative
- Negative
- None
Language Detection
Identify the language of a conversation so it can be transcribed, translated, or routed to the right agent or model. Detection works from either text or audio:
- Text-based: submit the text directly to the NLU service.
- Voice-based: call the language identification service, which determines the language from audio. See ASR Language Identifier for details.
Language Translation
Translate text between languages by submitting it to the NLU service. Specify the input and output languages explicitly, or enable automatic input-language detection and supply only the target language.
Neuron — Intent and Slot Extraction
Neuron is the platform’s intent and slot extraction service: determining what a speaker means. Using machine learning rather than keyword or rule matching, Neuron classifies each utterance to identify the speaker’s intent—the goal behind the request, such as Reset Password, Update Address, or Check Order Status—and extracts the slots (also called entities) that qualify it, such as a date, email address, product, location, or account number.
Because it interprets meaning rather than matching fixed phrases, Neuron stays accurate when customers speak naturally—varying their wording, using slang, switching languages, or supplying only part of the information. It carries context across conversational turns, so callers are not forced to repeat themselves, and it operates across both voice and digital channels.
For example, given the utterance “I need to check the status of my order from last week,” Neuron returns an intent of Check Order Status and a slot of date: last week—structured data an application can act on directly, routing the caller into the correct automation flow and bypassing rigid menu prompts.
For voice use cases, Neuron depends on an accurate transcription layer: the pipeline runs audio → text → meaning, with ASR & Transcription converting speech into the text Neuron interprets. Models are fully customized to each customer’s domain, terminology, and workflows, and can be developed for any language—backed by a linguistic team for intent and slot design, data preparation, and ongoing accuracy tuning.
Speaker Diarization
A single-channel (mono) recording of a multi-party conversation becomes a wall of undifferentiated text once transcribed—you know what was said, but not who said it. Speaker Diarization resolves “who spoke when,” taking mono audio with multiple speakers and segmenting it by speaker so each turn can be attributed correctly.
Accurate speaker attribution makes downstream analysis far more useful: agent-versus-caller talk time, turn-taking patterns, and per-speaker transcripts. Typical applications include transcription, meeting and call-center analytics, podcast and media production, and academic research.