Deepgram Applied Voice Learning Lab

Flux Conversation Observatory

See every turn, interruption, and response boundary.

Lab Live Solution Deliverables Capabilities

Synthetic fixtureDeepgram documentation verifiedManual validation required

Streaming text tells you what was heard. Flux turn events show when conversational state changes.

This workspace keeps provider messages, local lifecycle observations, threshold acknowledgements, and derived timing separate. It does not infer model reasoning or claim universal latency, accuracy, or production readiness.

Official contract checked 2026-07-28:Flux quickstart /v2/listen reference

Mode

Synthetic Replay

Connection

idle

Provider validation

synthetic only

Bounded memory

0/1500 events

Session & configuration

Choose evidence mode explicitly

Synthetic fixture — not a live Deepgram result

Uses the same normalizer, reducer, metrics, and export sanitizer as live mode. No network, microphone, or credential is used.

Replay scenario

An eager boundary is cancelled when the synthetic speaker continues after hesitating.

Flux configuration

Editable starting points

Assumption

Presets are hypotheses to test, never universal recommendations.

Preset

Balanced conversation starting point: A moderate starting point for ordinary two-party conversation; representative audio still decides fit. Exact values remain editable below and are never applied to a live session without the Apply action.

ModelEncodingSample rateTarget chunkeot_thresholdeager_eot_thresholdeot_timeout_msActual cadenceNot measured

Keyterms · up to 100Language hints · multilingual model only

Active: eot 0.7 · eager off · timeout 5000 ms. Failed updates never replace this acknowledged state.

Conversation timeline

Every received and locally observed boundary

Meaning first; sanitized payloads remain available for inspection.

0/0 events

No matching observations

Run a clearly labeled synthetic fixture, or explicitly prepare and start a live provider session. Nothing records automatically.

Turn intelligence

Objective session evidence

No model reasoning, accuracy guarantee, or universal benchmark is inferred.

Completed turns

0

Resumed turns

0

Unknown events

0

Config failures

0

Start → eager

Insufficient observations

n=0 · p95 requires 20 observations

Eager → end

Insufficient observations

n=0 · p95 requires 20 observations

Start → end

Insufficient observations

n=0 · p95 requires 20 observations

Observed chunk interval

Insufficient observations

n=0 · p95 requires 20 observations

Reconnect duration

Insufficient observations

n=0 · p95 requires 20 observations

Locally measured timing includes browser scheduling, capture, buffering, network, and transport effects. It is not a universal Deepgram benchmark.

Forced timeout: Insufficient provider evidence. The current TurnInfo wire contract does not expose an explicit forced-timeout reason, so this metric is not inferred.

Audio frames: 0 dropped · 0 delayed, when locally measurable.

Metric definitions

Completed turns: Count of turns for which an EndOfTurn provider event was received.
Resumed turns: Count of TurnResumed transitions observed after an eager boundary.
Start → eager: Local elapsed time from received StartOfTurn to received EagerEndOfTurn for the same completed turn.
Eager → end: Local elapsed time from received EagerEndOfTurn to received EndOfTurn for the same completed turn.
Start → end: Local elapsed time from received StartOfTurn to received EndOfTurn for the same completed turn.
Observed chunk interval: Elapsed browser time between locally emitted audio chunks. Scheduling and transport effects are included.
Reconnect duration: Local elapsed time from an explicit reconnect attempt to the next WebSocket open event.
Dropped frames: Frames intentionally not sent because the bounded WebSocket buffer exceeded the local safety limit.

Turn inspector

Received state sequence

0 grouped

No provider turn has been normalized yet.

Experimental orchestration demonstrator

Deterministic mock

idle

Select a turn or run a fixture to inspect the deterministic orchestration state.

No LLM, tool, TTS, or customer mutation runs. Duplicate tool actions cannot execute because the demonstrator has no execution adapter.

POC evidence generator

Sanitized scorecard

Qualitative reviewer notes - approved, redacted text only

Generate only after reviewing the run. Exports intentionally omit transcripts, raw audio, credentials, provider URLs, and internal stacks.

Architecture evidence

Validated Mermaid and typed handoffs

Generate from the active model, audio boundary, event path, configuration controls, and explicit optional-component labels.

Configuration laboratory

Requested, acknowledged, and rejected states remain distinct

0 transaction(s)

No configuration request has been sent in this bounded session. Editing does not change active settings.

Synthetic Replay is ready. No microphone, credential, network request, or provider usage starts automatically.