osborn 0.9.191 → 0.9.192

package/dist/config.d.ts

@@ -2,8 +2,8 @@ import type { McpServerConfig } from './claude-handler.js';
 export type VoiceMode = 'pipeline';
 export type EditMode = 'read-only' | 'edit';
 export type AgentMode = 'plan' | 'execute' | 'research';
-export type STTProvider = 'deepgram' | 'groq-whisper' | 'openai-whisper';
-export type TTSProvider = 'openai' | 'deepgram';
+export type STTProvider = 'soniox' | 'deepgram' | 'groq-whisper' | 'openai-whisper';
+export type TTSProvider = 'soniox' | 'openai' | 'deepgram';
 export interface DirectConfig {
     stt?: {
         provider?: STTProvider;

package/dist/config.js

@@ -47,13 +47,29 @@ const DEFAULT_CONFIG = {
     voiceMode: 'pipeline',
     direct: {
         stt: {
-            provider: 'deepgram',
-            model: 'nova-3',
+            // Soniox stt-rt-v4 — semantic endpointing (ML, not silence), word timestamps,
+            // custom vocabulary via context.terms. ~60% cheaper than nova-3 ($0.0017 vs $0.0043/min).
+            // Needs SONIOX_API_KEY.
+            provider: 'soniox',
+            model: 'stt-rt-v4',
+            // Previous: Deepgram nova-3, silence-based endpointing (550ms configured in voice-io.ts).
+            // Switch back: provider: 'deepgram', model: 'nova-3'
         },
         tts: {
-            provider: 'openai',
-            model: 'tts-1-hd',
-            voice: 'fable',
+            // Soniox tts-rt-v1 — real-time WebSocket streaming, speed control (0.7–1.3x),
+            // clean abort on interruption. Estimated ~$4–16/M chars vs OpenAI tts-1-hd $30/M.
+            // Pricing: $0.70/hr of generated speech (preview). Needs SONIOX_API_KEY.
+            provider: 'soniox',
+            model: 'tts-rt-v1',
+            voice: 'Maya',
+            // Previous: OpenAI tts-1-hd, voice fable — high quality, $30/M chars, ~500ms TTFB.
+            // Switch back: provider: 'openai', model: 'tts-1-hd', voice: 'fable'
+            // Other options already wired in voice-io.ts:
+            //   Rime Mist v3:     provider: 'rime',      voice: 'cove'  — 37ms TTFB, $30/M, WebSocket
+            //   Fish Audio s2-pro: provider: 'fishaudio', voice: '<id>'  — $15/M, voice cloning
+            //   Groq Orpheus:     provider: 'groq-orpheus', voice: 'autumn' — fast Groq chips, $22/M
+            //   Deepgram Aura-2:  provider: 'deepgram',  model: 'aura-2-asteria-en' — $15/M, ~100ms TTFB
+            //   OpenAI tts-1:     provider: 'openai',    model: 'tts-1', voice: 'fable' — $15/M
         },
     },
     mcpServers: {

package/dist/voice-io.d.ts

@@ -5,8 +5,9 @@
 import * as deepgram from '@livekit/agents-plugin-deepgram';
 import * as openai from '@livekit/agents-plugin-openai';
 import * as silero from '@livekit/agents-plugin-silero';
+import * as soniox from '@livekit/agents-plugin-soniox';
 export interface STTConfig {
-    provider: 'deepgram' | 'deepgram-flux' | 'groq-whisper' | 'openai-whisper';
+    provider: 'soniox' | 'deepgram' | 'deepgram-flux' | 'groq-whisper' | 'openai-whisper';
     model?: string;
     language?: string;
     /** Deepgram Flux: end-of-turn confidence threshold (0.0-1.0, default 0.7) */
@@ -15,7 +16,7 @@ export interface STTConfig {
     eotTimeoutMs?: number;
 }
 export interface TTSConfig {
-    provider: 'openai' | 'deepgram' | 'groq-orpheus' | 'fishaudio' | 'rime';
+    provider: 'soniox' | 'openai' | 'deepgram' | 'groq-orpheus' | 'fishaudio' | 'rime';
     voice?: string;
     model?: string;
 }
@@ -27,7 +28,7 @@ export interface VoiceIOConfig {
  * Create STT (Speech-to-Text) instance based on config
  * Note: Gemini STT is not available in Node.js, using Deepgram as default
  */
-export declare function createSTT(config: STTConfig): deepgram.STT | deepgram.STTv2 | openai.STT;
+export declare function createSTT(config: STTConfig): soniox.STT | deepgram.STT | deepgram.STTv2 | openai.STT;
 /**
  * Create TTS (Text-to-Speech) instance based on config
  */

package/dist/voice-io.js

@@ -7,13 +7,31 @@ import * as fishaudio from '@livekit/agents-plugin-fishaudio';
 import * as openai from '@livekit/agents-plugin-openai';
 import * as rime from '@livekit/agents-plugin-rime';
 import * as silero from '@livekit/agents-plugin-silero';
+import * as soniox from '@livekit/agents-plugin-soniox';
 /**
  * Create STT (Speech-to-Text) instance based on config
  * Note: Gemini STT is not available in Node.js, using Deepgram as default
  */
 export function createSTT(config) {
     switch (config.provider) {
+        case 'soniox':
+            // Soniox stt-rt-v4 — semantic endpointing: ML model holds on incomplete thoughts,
+            // commits on natural sentence ends. maxEndpointDelayMs 500–3000ms (minimum = fastest).
+            // endpointLatencyAdjustmentLevel 0–3: higher = more aggressive latency reduction
+            // while keeping semantic smarts. context.terms biases recognition toward code vocab.
+            return new soniox.STT({
+                model: (config.model || 'stt-rt-v4'),
+                languageHints: config.language ? [config.language] : ['en'],
+                maxEndpointDelayMs: 1200, // give model room to decide on mid-thought pauses
+                endpointLatencyAdjustmentLevel: 2, // aggressive but not max — good for voice assistant
+                context: {
+                    terms: ['Claude', 'TypeScript', 'LiveKit', 'Deepgram', 'npm', 'Railway', 'Fly.io'],
+                },
+            });
         case 'deepgram':
+            // Previous default. Silence-based endpointing (550ms configured = wait for 550ms quiet).
+            // Fast and reliable, but commits on any pause — doesn't understand mid-thought hesitations.
+            // Switch back: provider: 'deepgram', model: 'nova-3'
             return new deepgram.STT({
                 model: (config.model || 'nova-3'),
                 language: config.language || 'en',
@@ -49,13 +67,30 @@ export function createSTT(config) {
 export function createTTS(config) {
     let tts;
     switch (config.provider) {
+        case 'soniox':
+            // Soniox tts-rt-v1 — real-time WebSocket streaming, clean abort on interruption.
+            // Estimated ~$4–16/M chars ($0.70/hr of generated speech, preview pricing).
+            // speed: 0.7–1.3x. voices: Maya (female), others at soniox.com/docs/tts/voices.
+            // Previous TTS: OpenAI tts-1-hd (fable) — $30/M chars, ~500ms TTFB, HTTP chunked.
+            // Switch back: provider: 'openai', model: 'tts-1-hd', voice: 'fable'
+            tts = new soniox.TTS({
+                model: (config.model || 'tts-rt-v1'),
+                voice: config.voice || 'Maya',
+                speed: 1.0,
+            });
+            break;
         case 'openai':
+            // tts-1-hd: $30/M chars, ~500ms TTFB, 6 voices: alloy echo fable onyx nova shimmer.
+            // tts-1 (cheaper): $15/M chars, slightly lower quality, same voices.
             tts = new openai.TTS({
                 voice: config.voice || 'alloy',
                 model: config.model || 'tts-1',
             });
             break;
         case 'deepgram':
+            // Aura-2 voices: aura-2-asteria-en, aura-2-luna-en, aura-2-stella-en, aura-2-hera-en
+            // aura-2-orion-en, aura-2-arcas-en, aura-2-perseus-en, aura-2-angus-en, aura-2-orpheus-en
+            // ~$15/M chars, ~100ms TTFB.
             tts = new deepgram.TTS({
                 model: (config.model || 'aura-2-asteria-en'),
             });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.191",
+  "version": "0.9.192",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {
@@ -39,6 +39,7 @@
     "@livekit/agents-plugin-openai": "1.4.6",
     "@livekit/agents-plugin-rime": "1.4.6",
     "@livekit/agents-plugin-silero": "1.4.6",
+    "@livekit/agents-plugin-soniox": "^1.8.1",
     "@livekit/rtc-node": "0.13.29",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@smithery/api": "^0.48.0",