Docs /tutorials/ Conversation Intelligence
tutorials / use-cases

Conversation Intelligence

Transcription, speaker diarization, sentiment and analysis of spoken conversations.

audioasrdiarizationsentimenttopic-modelingDocumentation latest

Source: Conversation Intelligence — GitHub

Conversation Intelligence is a voice-analysis demonstration built with Sinapsis. It transcribes uploaded audio, labels the speakers, analyses sentiment, produces an LLM-based conversation assessment, and offers recommended follow-up actions.

The implementation is in sinapsis-demos/conversation-intelligence; all audio-processing code, model artifacts, and YAML configurations remain there.

What it demonstrates

  • Speech transcription and speaker diarization using WhisperxDiarizedTranscription.

  • Speaker-aware sentiment processing with a Transformers-based financial sentiment model.

  • Structured LLM analysis of the resulting transcript.

  • Suggested actions from a contextual-bandit Agent.

  • Topic-model visualisation backed by BERTopic artifacts.

The Gradio application orchestrates multiple distinct Sinapsis Agents. This is an application workflow, not one declarative Agent graph; the internal Template connections of each Agent are defined separately in their YAML files.

Application flow

◇ DiagramDrag nodes · Scroll to zoom · Drag canvas to pan
flowchart TD
    AUDIO([Uploaded audio]) --> RESAMPLE[Audio preprocessing<br/>16 kHz when needed]
    RESAMPLE --> ASR[ASR and diarization Agent<br/>WhisperxDiarizedTranscription]
    ASR --> TRANSCRIPT[Speaker-labelled transcript]
    TRANSCRIPT --> SENT[Sentiment Agent<br/>PatternAggregatorSplitter + Transformer]
    TRANSCRIPT --> LLM[Structured LLM analysis Agent]
    LLM --> SUMMARY[Conversation summary]
    SUMMARY --> ACTION[Contextual-bandit Agent]
    MODEL[(Saved BERTopic visualization)] --> VIEW[Topic display]
    SENT --> UI([Gradio results])
    LLM --> UI
    ACTION --> UI
    VIEW --> UI

Green-bordered nodes reference Sinapsis templates. Hover or focus them to fetch documentation.

View Mermaid source
mermaid
flowchart TD
    AUDIO([Uploaded audio]) --> RESAMPLE[Audio preprocessing<br/>16 kHz when needed]
    RESAMPLE --> ASR[ASR and diarization Agent<br/>WhisperxDiarizedTranscription]
    ASR --> TRANSCRIPT[Speaker-labelled transcript]
    TRANSCRIPT --> SENT[Sentiment Agent<br/>PatternAggregatorSplitter + Transformer]
    TRANSCRIPT --> LLM[Structured LLM analysis Agent]
    LLM --> SUMMARY[Conversation summary]
    SUMMARY --> ACTION[Contextual-bandit Agent]
    MODEL[(Saved BERTopic visualization)] --> VIEW[Topic display]
    SENT --> UI([Gradio results])
    LLM --> UI
    ACTION --> UI
    VIEW --> UI

The ASR Agent returns audio annotations containing diarized transcript segments. The application writes a speaker-labelled transcript to a text file and passes that file to the subsequent sentiment and LLM Agents. A separate contextual-bandit Agent uses the generated summary to propose next actions.

Topic-visualisation distinction: In the supplied webapp.py, the call that would run bertopic_agent is commented out. The UI instead reads an existing artifacts/bertopic_prediction_visualization.html if present. The BERTopic training, prediction, and visualisation YAML files are provided separately, but the live interface does not currently execute all those steps.

Source files

FilePurpose
README.mdOverview and development setup.
webapp.pyApplication orchestration and Gradio interface.
configs/speech_to_text.yamlWhisperX ASR and diarization.
configs/sentiment_analysis.yamlSpeaker-aware sentiment analysis.
configs/llm_structured.yamlStructured LLM analysis.
configs/contextual_bandits.yamlContextual-bandit action proposals.
configs/bertopic_predict.yamlBERTopic prediction configuration.
configs/bertopic_fit_model.yamlExample BERTopic model training.
configs/visualize_preds.yamlSaved-topic visualisation.
docker/DockerfileGPU runtime image.
docker/compose_test.yamlDevelopment Compose configuration.

The demo also contains model artifacts, prompt files, and development scripts. They remain in the original repository; this page does not duplicate model files or test data. The source links currently target main and should be pinned for release.

Requirements and running the demo

The README specifies Docker/Compose with BuildKit, an NVIDIA GPU and NVIDIA Container Toolkit (approximately 8 GB or more GPU memory for ASR/LLM tasks), and an available sinapsis-nvidia:base image. The application exposes a Gradio interface on port 7860.

From the demos repository root, the intended command is:

Terminalbash
cd conversation-intelligence/docker
docker compose -f compose_test.yaml up --build

The command has not been executed or verified in a clean environment. Some build dependencies rely on private Git access.

To understand why the individual Agents can remain separately configured, see Agent Model and DataContainers.

Published from Git revision 7b7ba25d627b