osborn 0.9.192 → 0.9.194

package/dist/config.js

@@ -61,7 +61,7 @@ const DEFAULT_CONFIG = {
             // Pricing: $0.70/hr of generated speech (preview). Needs SONIOX_API_KEY.
             provider: 'soniox',
             model: 'tts-rt-v1',
-            voice: 'Maya',
+            voice: 'Victoria',
             // 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:
@@ -210,16 +210,19 @@ export function getVoiceMode(_config) {
 export function getDirectConfig(config) {
     const defaults = DEFAULT_CONFIG.direct;
     const userConfig = config.direct || {};
+    // Env vars take priority over config.yaml — escape hatch for stale volume configs.
+    // Set OSBORN_STT_PROVIDER / OSBORN_TTS_PROVIDER / OSBORN_TTS_VOICE on the machine
+    // to override whatever is baked into /workspace/.osborn/config.yaml.
     return {
         stt: {
-            provider: userConfig.stt?.provider || defaults.stt.provider,
-            model: userConfig.stt?.model || defaults.stt.model,
+            provider: (process.env.OSBORN_STT_PROVIDER || userConfig.stt?.provider || defaults.stt.provider),
+            model: process.env.OSBORN_STT_MODEL || userConfig.stt?.model || defaults.stt.model,
             language: userConfig.stt?.language || 'en',
         },
         tts: {
-            provider: userConfig.tts?.provider || defaults.tts.provider,
-            model: userConfig.tts?.model || defaults.tts.model,
-            voice: userConfig.tts?.voice || defaults.tts.voice,
+            provider: (process.env.OSBORN_TTS_PROVIDER || userConfig.tts?.provider || defaults.tts.provider),
+            model: process.env.OSBORN_TTS_MODEL || userConfig.tts?.model || defaults.tts.model,
+            voice: process.env.OSBORN_TTS_VOICE || userConfig.tts?.voice || defaults.tts.voice,
         },
     };
 }

package/dist/index.js

@@ -30,7 +30,7 @@ const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 import { createPatch } from 'diff';
 import { loadConfig, getMcpServers, getEnabledMcpServerNames, getVoiceMode, getDirectConfig, listSessions, listAllClaudeSessions, invalidateSessionListCache, getMostRecentSessionId, sessionExists, getSessionSummary, getConversationHistory, ensureSessionWorkspace, getMcpServerStatusList, buildMcpServersForKeys, listWorkspaceArtifacts } from './config.js';
-import { createSTT, createTTS, DIRECT_MODE_STT, DIRECT_MODE_TTS } from './voice-io.js';
+import { createSTT, createTTS, DIRECT_MODE_TTS } from './voice-io.js';
 import { createClaudeLLM, NAMED_AGENTS, applyTurbo } from './claude-llm.js';
 import { clearPipelineFastBrainSession, prewarmBM25Index } from './pipeline-fastbrain.js';
 import { getIndexPath, buildSummaryIndex } from './summary-index.js';
@@ -3173,8 +3173,8 @@ async function main() {
     // Create DIRECT session (STT + Claude Agent SDK + TTS)
     async function createDirectSession(resumeSessionId, llmOverride) {
         console.log('🎯 Creating direct session...');
-        const stt = createSTT(DIRECT_MODE_STT);
-        const tts = createTTS(DIRECT_MODE_TTS);
+        const stt = createSTT(directConfig.stt);
+        const tts = createTTS(directConfig.tts);
         // Create Claude LLM wrapper — direct mode uses speech-optimized system prompt
         // skipTTSQueue: bypass LiveKit's BufferedTokenStream, use session.say() instead
         // llmOverride: pipeline mode passes PipelineDirectLLM which wraps its own ClaudeLLM

package/dist/voice-io.js

@@ -162,13 +162,15 @@ export const DIRECT_MODE_STT = {
     // provider: 'groq-whisper', model: 'whisper-large-v3-turbo',          // Batch — needs VAD
     // provider: 'openai-whisper', model: 'whisper-1',                     // Batch — needs VAD
     // provider: 'deepgram-flux', model: 'flux-general-en', language: 'en',  // Streaming, ML-based turn detection — Flux V2 has silent keepalive timeout bug (~30s silence kills connection)
-    provider: 'deepgram', model: 'nova-3', language: 'en', // Streaming, silence-based endpointing (550ms)
+    // provider: 'deepgram', model: 'nova-3', language: 'en',  // Silence-based endpointing (550ms) — $0.0043/min
+    provider: 'soniox', model: 'stt-rt-v4', language: 'en', // Semantic endpointing, ~60% cheaper ($0.0017/min) — needs SONIOX_API_KEY
 };
 export const DIRECT_MODE_TTS = {
     // provider: 'deepgram', model: 'aura-2-asteria-en',  // WebSocket-based: handles TTS abort cleanly — but quality rejected (run-on sentences)
     // provider: 'openai', model: 'tts-1', voice: 'fable',  // HTTP streaming: throws APIUserAbortError on interrupt → unrecoverable session crash
-    provider: 'openai', model: 'tts-1-hd', voice: 'fable', // 0.9.70: test tts-1-hd — tts-1 had chronic per-sentence HTTP hangs (40s SDK watchdog → APIUserAbortError mid-message)
+    // provider: 'openai', model: 'tts-1-hd', voice: 'fable',  // $30/M chars, ~500ms TTFB — fallback if Soniox has issues
     // provider: 'groq-orpheus', model: 'canopylabs/orpheus-v1-english', voice: 'autumn',  // $22/M chars — voices: autumn, diana, hannah, austin, daniel, troy
     // provider: 'fishaudio', model: 's2-pro', voice: '<voice-id>',  // $15/M chars — blind test winner, HTTP streaming (test abort behavior)
     // provider: 'rime', model: 'mistv3', voice: 'cove',  // $30/M chars, 37ms TTFB — WebSocket (safe on interruption); voices: aurora, ember, cove
+    provider: 'soniox', model: 'tts-rt-v1', voice: 'Victoria', // WebSocket streaming, ~$4–16/M chars, British female — needs SONIOX_API_KEY
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.192",
+  "version": "0.9.194",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {