diff --git a/CHANGELOG.md b/CHANGELOG.md
index 74f79ab..eb449c2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,17 @@ All notable changes to this project are documented here. The format is based on
 
 ## [Unreleased]
 
+### Changed
+
+- **BREAKING: Clip + Voice contracts aligned to the LIVE gateway** (verified by probing
+  `api.wave.online`). SDK/CLI clients generated from the previous spec will break:
+  - `ClipCreate` now requires `source` (recording id string) + `in`/`out` relative time strings
+    (`5s`, `2m`), replacing the rejected `videoId`/`startTime`/`endTime` numeric shape.
+  - `POST /voice/generate` → `POST /voice` (the live path); `VoiceGenerateRequest` requires only
+    `text` (`voiceId` optional, also sent as `voice_id` by the SDK). The 200 response now documents
+    that the primary path returns raw `audio/mpeg` bytes directly. The old `/voice/generate` path
+    is not kept as a deprecated alias because the live gateway does not serve it.
+
 ### Added
 
 - **MoQ join-token mint surface** (`openapi.yaml`) — the Media over QUIC product had no spec at
diff --git a/openapi.yaml b/openapi.yaml
index dc10210..08c608e 100644
--- a/openapi.yaml
+++ b/openapi.yaml
@@ -390,7 +390,7 @@ paths:
                     items:
                       $ref: '#/components/schemas/Voice'
 
-  /voice/generate:
+  /voice:
     post:
       tags: [Voice]
       summary: Generate speech from text
@@ -404,20 +404,22 @@ paths:
       responses:
         '200':
           description: >
-            Speech generated. The gateway returns one of three shapes depending on
-            the engine path: an inline JSON payload with base64 audio + character
-            `alignment` (single round-trip, carries word timestamps), an async job
-            to poll, or raw audio bytes when timestamps were not requested.
+            Speech generated. Returns the raw MP3 bytes (`audio/mpeg`) by default (verified
+            against the live gateway: POST `/voice` with `{ text }` returns the MP3 directly).
+            The `application/json` shapes are returned only when the request opts out of the
+            default: requests that ask for character timestamps receive a `VoiceSynthesisInline`
+            payload (base64 audio + `alignment`), and engines that process asynchronously return
+            a `VoiceGeneration` job to poll.
           content:
+            audio/mpeg:
+              schema:
+                type: string
+                format: binary
             application/json:
               schema:
                 oneOf:
                   - $ref: '#/components/schemas/VoiceSynthesisInline'
                   - $ref: '#/components/schemas/VoiceGeneration'
-            audio/mpeg:
-              schema:
-                type: string
-                format: binary
         '401':
           $ref: '#/components/responses/Unauthorized'
         '403':
@@ -2238,14 +2240,17 @@ components:
 
     ClipCreate:
       type: object
-      required: [videoId, startTime, endTime]
+      required: [source, in, out]
       properties:
-        videoId:
+        source:
           type: string
-        startTime:
-          type: number
-        endTime:
-          type: number
+          description: Recording id the clip is cut from (e.g. `rec_abc123`). Verified against the live gateway — the older `{ type, id, start_time, end_time }` source object is rejected.
+        in:
+          type: string
+          description: Start offset as a relative time string, e.g. `5s` or `2m`.
+        out:
+          type: string
+          description: End offset as a relative time string, e.g. `10s` or `1m30s`.
         title:
           type: string
         description:
@@ -2398,12 +2403,20 @@ components:
 
     VoiceGenerateRequest:
       type: object
-      required: [voiceId, text]
+      required: [text]
+      # `voiceId` and `voice_id` are wire aliases for the same setting; supplying both is invalid.
+      not:
+        required: [voiceId, voice_id]
       properties:
+        text:
+          type: string
+          description: Text to convert to speech.
         voiceId:
           type: string
-        text:
+          description: Voice id to use. Optional; the gateway picks a default when omitted. Also accepted on the wire as the `voice_id` alias.
+        voice_id:
           type: string
+          description: Snake_case wire alias for `voiceId` (the form the SDK sends). Provide either `voiceId` or `voice_id`, not both.
         stability:
           type: number
           minimum: 0