Add TTS backends (ElevenLabs, Kokoro) and /relay/tts

2026-06-13 13:36:05 -05:00
parent 0aa648706e
commit 04dcf86fa4
3 changed files with 538 additions and 0 deletions
@@ -0,0 +1,109 @@
+// ElevenLabs TTS backend — the cloud alternative to operator-hardware
+// Kokoro, mirroring how the transcribe/analyze pipelines treat Gemini
+// as a swappable cloud fallback to the Parakeet/vLLM hardware path.
+//
+// Selected when relay_tts_backend_preference routes here (or when
+// Kokoro is unavailable and a fallback is allowed) AND
+// relay_elevenlabs_api_key is configured. Like Kokoro, ElevenLabs
+// handles full multi-sentence paragraphs in one call — one request per
+// topic summary, no chunking.
+//
+// NOTE: This path is implemented to ElevenLabs' documented API but is
+// UNTESTED against a live key (the operator hadn't supplied one at
+// build time). The Kokoro path is the tested default. Before relying on
+// ElevenLabs in production, set relay_elevenlabs_api_key +
+// relay_elevenlabs_voice_id and smoke-test one /relay/tts call.
+//
+// Output: { audio: Buffer, contentType: "audio/mpeg", durationSeconds,
+//           voice, model } — durationSeconds is null (we don't decode
+//           the MP3 frame count here; the Recap server measures it when
+//           transcoding/caching).
+
+const API_BASE = "https://api.elevenlabs.io/v1";
+const DEFAULT_MODEL = "eleven_turbo_v2_5";
+const DEFAULT_TIMEOUT_MS = 120_000;
+
+export function createElevenLabsBackend({
+  apiKey = "",
+  voiceId = "",
+  model = DEFAULT_MODEL,
+  timeoutMs = DEFAULT_TIMEOUT_MS,
+} = {}) {
+  const configured = !!(apiKey && voiceId);
+
+  return {
+    hasTts: configured,
+    kind: "elevenlabs",
+
+    async synthesize({ text, voice }) {
+      if (!apiKey) {
+        const e = new Error(
+          "ElevenLabs TTS is not configured — set relay_elevenlabs_api_key"
+        );
+        e.status = 503;
+        throw e;
+      }
+      // `voice` from the client overrides the operator default voice id
+      // when present (the Recap client may let a user pick a voice).
+      const chosenVoice = (voice || voiceId || "").trim();
+      if (!chosenVoice) {
+        const e = new Error(
+          "ElevenLabs TTS has no voice — set relay_elevenlabs_voice_id"
+        );
+        e.status = 503;
+        throw e;
+      }
+      const cleaned = (text || "").replace(/\s+/g, " ").trim();
+      if (!cleaned) {
+        const e = new Error("TTS input text is empty");
+        e.status = 400;
+        throw e;
+      }
+
+      const url = `${API_BASE}/text-to-speech/${encodeURIComponent(chosenVoice)}`;
+      let res;
+      try {
+        // Public-internet call — use the global fetch with full cert
+        // validation (NOT lanFetch, which is scoped to LAN/Spark Control).
+        res = await fetch(url, {
+          method: "POST",
+          headers: {
+            "xi-api-key": apiKey,
+            "Content-Type": "application/json",
+            Accept: "audio/mpeg",
+          },
+          body: JSON.stringify({
+            text: cleaned,
+            model_id: model || DEFAULT_MODEL,
+          }),
+          signal: AbortSignal.timeout(timeoutMs),
+        });
+      } catch (err) {
+        const e = new Error(`ElevenLabs TTS network error: ${err?.message || err}`);
+        e.status = 502;
+        throw e;
+      }
+      if (!res.ok) {
+        let body = "";
+        try {
+          body = await res.text();
+        } catch {}
+        const e = new Error(
+          `ElevenLabs TTS ${res.status}: ${body.slice(0, 300)}`
+        );
+        e.status = res.status;
+        throw e;
+      }
+      const audio = Buffer.from(await res.arrayBuffer());
+      return {
+        audio,
+        contentType: "audio/mpeg",
+        durationSeconds: null,
+        sentenceCount: null,
+        attempts: 1,
+        voice: chosenVoice,
+        model: model || DEFAULT_MODEL,
+      };
+    },
+  };
+}