Voice Feature Parity Implementation Plan¶

Full voice calling feature parity for omniagent+omnivoice with openclaw.

Status¶

Phase	Module	Package	Status
1	omnivoice-core	`storage/`	Completed
1	omnivoice-core	`bargein/`	Completed
2	omni-openai	`omnivoice/realtime/`	Completed
2	omni-google	`omnivoice/`	Completed
3	omniskill	`voicetools/`	Completed

Summary¶

Feature	Module	Location
OpenAI Realtime API	omni-openai	`omnivoice/realtime/`
Gemini Live API	omni-google	`omnivoice/`
Barge-in Detection	omnivoice-core	`bargein/`
Call State Persistence	omnivoice-core	`storage/`
Agent Consult Tools	omniskill	`voicetools/`

Phase 1: Core Infrastructure - Completed¶

omnivoice-core/storage/¶

Session state persistence with pluggable backends:

type SessionStore interface {
    Save(ctx context.Context, session *SessionState) error
    Load(ctx context.Context, sessionID string) (*SessionState, error)
    Delete(ctx context.Context, sessionID string) error
    ListActive(ctx context.Context) ([]string, error)
    UpdateHeartbeat(ctx context.Context, sessionID string) error
    Close() error
}

Implementations:

MemoryStore - In-memory for development/testing
RedisStore - Redis-backed for production multi-instance

Files:

store.go - Interface definitions
types.go - SessionState, Turn, SessionMetrics
memory.go - In-memory implementation
redis.go - Redis implementation

omnivoice-core/bargein/¶

Barge-in detection for natural interruption handling:

type Detector struct {
    config        Config
    ttsPipeline   interface{ IsActive() bool; Stop() }
    onInterrupt   func(Event)
}

func (d *Detector) AttachTTS(p interface{ IsActive() bool; Stop() })
func (d *Detector) AttachSTTEvents(events <-chan STTEvent)
func (d *Detector) OnInterrupt(handler func(Event))

Interruption Modes:

ModeImmediate - Interrupt as soon as user speech detected
ModeAfterSentence - Wait for natural pause
ModeDisabled - Ignore user speech during agent output

Phase 2: Realtime APIs - Completed¶

omni-openai/omnivoice/realtime/¶

OpenAI Realtime API WebSocket client for native voice-to-voice (~100ms latency):

type RealtimeProvider struct {
    client *Client
    config Config
}

func (p *RealtimeProvider) ProcessAudioStream(
    ctx context.Context,
    audioIn <-chan []byte,
    config ProcessConfig,
) (<-chan AudioChunk, <-chan Transcript, error)

Audio Format:

Input: PCM16 24kHz mono
Output: PCM16 24kHz mono

Features:

Session management via WebSocket
Function calling support
Voice Activity Detection (VAD)
11 voice options

omni-google/omnivoice/¶

Gemini Live API WebSocket client for real-time voice-to-voice:

type RealtimeProvider struct {
    apiKey string
    config Config
}

func (p *RealtimeProvider) ProcessAudioStream(
    ctx context.Context,
    audioIn <-chan []byte,
    config ProcessConfig,
) (<-chan AudioChunk, <-chan Transcript, error)

Audio Format:

Input: PCM16 16kHz mono
Output: PCM16 24kHz mono

Features:

Bidirectional WebSocket streaming
Function calling support
Google Search grounding
Code execution
5 voice options

Phase 3: Voice Tools - Completed¶

omniskill/voicetools/¶

AI agent tools for voice call control:

func NewVoiceSkill(callCtx CallContext) *skill.Skill {
    return skill.New("voice_control",
        skill.WithTool(NewTransferCallTool(callCtx)),
        skill.WithTool(NewHoldCallTool(callCtx)),
        skill.WithTool(NewUnholdCallTool(callCtx)),
        skill.WithTool(NewConsultAgentTool(callCtx)),
        skill.WithTool(NewConferenceTool(callCtx)),
    )
}

Tools:

Tool	Description
`transfer_call`	Transfer to number or queue
`hold_call`	Place caller on hold
`unhold_call`	Resume from hold
`consult_agent`	Query specialist without transfer
`add_to_conference`	Add participants

Verification¶

All packages pass tests:

cd ~/go/src/github.com/plexusone/omnivoice-core && go test ./storage/... ./bargein/...
cd ~/go/src/github.com/plexusone/omni-openai && go test ./omnivoice/realtime/...
cd ~/go/src/github.com/plexusone/omni-google && go test ./omnivoice/...
cd ~/go/src/github.com/plexusone/omniskill && go test ./voicetools/...

All packages pass linting:

golangci-lint run

Dependencies Added¶

Module	Dependency	Purpose
omnivoice-core	`github.com/redis/go-redis/v9`	Redis session store
omni-openai	`github.com/gorilla/websocket`	WebSocket client
omni-google	`github.com/gorilla/websocket`	WebSocket client

Environment Variables¶

# OpenAI Realtime
OPENAI_API_KEY=sk-...

# Gemini Live
GOOGLE_API_KEY=...
# or GEMINI_API_KEY=...

# Redis (optional, for persistence)
REDIS_URL=redis://localhost:6379

Documentation¶

Updated documentation for each module:

Module	README	MkDocs Pages
omnivoice-core	Updated with storage/, bargein/	storage.md, bargein.md
omni-openai	Updated with realtime/	providers/realtime.md
omni-google	Updated with omnivoice/	omnivoice/index.md
omniskill	Updated with voicetools/	voicetools/index.md