OpenRouter media generation: ElevenLabs TTS, batch STT, video, images and music

All of these worked with a normal OpenRouter API key; no separate ElevenLabs or Google plan was needed. Verified 18 October 2026.

ElevenLabs text-to-speech

• POST https://openrouter.ai/api/v1/audio/speech, JSON {model, input, voice, response_format: "mp3"}. The response body is raw audio bytes, not JSON. The default format is pcm, so request mp3 explicitly.
• Models: elevenlabs/eleven-v4, eleven-v4-turbo, eleven-v3, eleven-v3-conversational, eleven-multilingual-v2, eleven-flash-v2.5, and others. v3/v4 cost $0.00004 per character, so a 30-second script costs about 1 cent.
• Voices are names only: george, sarah, adam, alice, bella, bill, brian, callum, charlie, chris, daniel, eric, harry, jessica, laura, liam, lily, matilda, river, roger, will. There is no New Zealand voice; charlie is Australian.
• v3/v4 audio tags such as [New Zealand accent], [excited] and [polite cough] are honoured and not read aloud (checked by transcription).
• The text "Arr" is read as a pirate growl; "Ah" gives the intended pronunciation.
• To check a clip without listening, send it as input_audio (base64 mp3) to google/gemini-3.8-flash and ask for a transcript. Its accent guesses are unreliable.
• Streaming: response_format: "pcm" returns audio/pcm;rate=24000;channels=1 with chunked transfer, so it can be played as it arrives. First byte measured from New Zealand on 18 October 2026: about 1.4-1.6 s on a new connection and 0.5-0.75 s with a reused connection (eleven-flash-v2.5, eleven-v4-turbo). Direct ElevenLabs Flash was about 0.1-0.2 s, so OpenRouter TTS is only a fallback for real-time voice.

Speech-to-text (batch only)

• POST https://openrouter.ai/api/v1/audio/transcriptions, JSON {model, input_audio: {data: <base64>, format: "wav"}} returns {"text": ..., "usage": {seconds|tokens, cost}}. Verified with a 2.9 s WAV clip.
• Working models: deepgram/nova-3 (1.26 s round trip, about $0.0043/min), openai/gpt-4o-mini-transcribe (0.61 s, $0.00011 for the clip), openai/whisper-1 (2.8 s, $0.006/min). elevenlabs/scribe-v1 does not exist there.
• There is no streaming or WebSocket STT: no interim words, VAD or end-of-turn events. A real-time voice agent therefore needs its own VAD and pays the transcription round trip on every turn.

Image-to-video

• POST /api/v1/videos with {model, prompt, duration, aspect_ratio: "9:16", resolution: "720p", generate_audio: false, frame_images: [{type: "image_url", image_url: {url}, frame_type: "first_frame"}]}. Base64 data URLs work for the frame image.
• The POST returns {id, polling_url, status}. Poll GET /api/v1/videos/{id} until completed or failed, then download with GET /api/v1/videos/{id}/content?index=0.
• GET /api/v1/videos/models lists every model's durations, ratios and pricing_skus.
• google/veo-3.1-lite at 720p without audio costs $0.03 per second, with good quality for a 3D cartoon style. Durations are 4, 6 or 8 seconds only.
• In multi-character shots Veo can morph characters into generic ghosts. Adding "keep exact appearance, single continuous shot, characters do not transform" to the prompt fixed it.

Images

• Chat completions with modalities: ["image","text"] and image_config: {aspect_ratio: "9:16"}. Reference images passed as image_url parts keep characters consistent.
• google/gemini-nano-banana-2.1 often returned 429 because of the shared upstream pool, and parallel calls made that worse. google/gemini-3.1-flash-image (about $0.07 per image) worked when called one at a time with a 429 retry.

Music

• google/lyria-3-clip-preview costs $0.04 per 30-second instrumental clip. Use chat completions with modalities: ["audio","text"]; it requires stream: true. Concatenating the base64 delta.audio.data chunks gives an mp3.

Budget example

One 31-second vertical video (6 keyframes, 6 Veo clips with one redo, voices, music) cost about $1.84.