chat: mirror live librechat.yaml — STT/TTS endpoints, voxtral models, seeded speechTab, STT language default 'en'

Syncs the repo to the deployed /app/data/librechat.yaml (md5 a856831404d5b522cd2e47a8e63466ad,
verified in-container 2026-11-07): full /v1/audio/* endpoint URLs, tinfoil/voxtral-small-24b
(STT) + tinfoil/voxtral-tts (TTS) with the 20 provider voices, seeded speechTab client
defaults (external engines), and languageSTT 'en' pinned as the transcription-language
default (members' own saved settings always win) after a first-press Spanish misdetect.
This run's TTS latency fix lives in the gateway (code/litellm 5d38403), not here.

librechat.yaml

@@ -133,27 +133,48 @@ interface:
           rgb-border-light: '74 82 64'
 
 
-# Speech (STT/TTS) via the co-op gateway. Uses a dedicated LiteLLM virtual key
+# Speech (STT/TTS) via the co-op gateway. LibreChat POSTs to `url` VERBATIM
-# (alias: librechat-speech, $5/30d budget, restricted to the audio models) so
+# (no path is appended), so the FULL endpoint paths are required here.
-# speech never runs on the gateway master key. The gateway is a public
+# Uses a dedicated LiteLLM virtual key (alias: librechat-speech, $5/30d
+# budget, restricted to the audio models) injected as STT_API_KEY/TTS_API_KEY
+# app env — speech never runs on the master key. The gateway is a public
 # endpoint, so no SSRF allowedAddresses exemption is needed.
+# STT: tinfoil/voxtral-small-24b — token-priced (~$0.00008/typical dictation
+# request on the gateway) vs the flat $0.01/request of
+# tinfoil/whisper-large-v3-turbo; same TEE provider, equal quality on short
+# clips (verified round-trip 2026-11-07). Whisper stays allowed on the key as
+# a one-line fallback swap. TTS: tinfoil/voxtral-tts, $0.002/request at cost;
+# voices are the provider's supported speakers (an unknown name 400s).
 speech:
   tts:
     openai:
-      url: "https://gateway.inference.coop/v1"
+      url: "https://gateway.inference.coop/v1/audio/speech"
-      apiKey: "${TTS_API_KEY}"
+      apiKey: ${TTS_API_KEY}
       model: "tinfoil/voxtral-tts"
-      voices: ["casual_female"]
+      voices: ["ar_male", "casual_female", "casual_male", "cheerful_female", "de_female", "de_male", "es_female", "es_male", "fr_female", "fr_male", "hi_female", "hi_male", "it_female", "it_male", "neutral_female", "neutral_male", "nl_female", "nl_male", "pt_female", "pt_male"]
   stt:
     openai:
-      url: "https://gateway.inference.coop/v1"
+      url: "https://gateway.inference.coop/v1/audio/transcriptions"
-      apiKey: "${STT_API_KEY}"
+      apiKey: ${STT_API_KEY}
-      model: "tinfoil/whisper-large-v3-turbo"
+      model: "tinfoil/voxtral-small-24b"
+  # Seeded client defaults (Settings -> Speech): members who never saved
+  # their own speech settings get the External engines out of the box;
+  # a member's own saved settings always win. "Automatic playback" stays
+  # opt-in per member (set textToSpeech.automaticPlayback: true to change).
   speechTab:
     conversationMode: false
     advancedMode: false
-    speechToText: true
+    speechToText:
-    textToSpeech: true
+      engineSTT: external
+      # Pin the transcription LANGUAGE default to English. Without it the
+      # client sends no language param and voxtral-small-24b auto-detects —
+      # which misfired once as Spanish on a short first mic press (reported
+      # 2026-11-07). Members who save their own language in Settings -> Speech
+      # always override this default.
+      languageSTT: "en"
+    textToSpeech:
+      engineTTS: external
+      voice: casual_female
 
 # Web search: self-hosted SearXNG (search.inference.coop, keyless JSON API)
 # for search, keyless Keenable for page scraping, no reranker. The instance