Text to Speech

Convert text to speech through the rebyte TTS endpoint. One endpoint, two providers, one voice catalog.

Requires Rebyte API auth — $AUTHTOKEN and $APIURL are set up per the agent's system prompt.

Picking a voice

Every voice has an explicit provider prefix. Pick the voice that matches the tone you want, then compose style per its column below.

OpenAI voices (style via instructions field)

Uses gpt-audio-mini by default (fast, cheap). Pass model: "gpt-audio" for highest quality. Style is controlled via the instructions field — a natural language directive for delivery.

Additional voices: openai:nova, openai:alloy, openai:echo, openai:fable, openai:onyx, openai:shimmer.

Gemini voices (style via natural language in text)

Backend: gemini-3.1-flash-tts-preview. Style is controlled by writing natural-language directions as part of the text itself. The model interprets the directions and speaks accordingly. Think of it like directing an actor — describe how to deliver the line, then give the line.

Additional Gemini voices: gemini:Orus, gemini:Zephyr, gemini:Callirrhoe, gemini:Autonoe, gemini:Enceladus, gemini:Iapetus, gemini:Umbriel, gemini:Algieba, gemini:Despina, gemini:Algenib, gemini:Rasalgethi, gemini:Achernar, gemini:Schedar, gemini:Gacrux, gemini:Sulafat.

Style direction patterns (natural language — these are examples, not an enum):

• Say in a whisper: We have to be quiet here.
• Say excitedly and fast: You are not going to believe what just happened!
• Say sadly, slowly: I don't think this is going to work out.
• Deadpan delivery: Yes. That is how physics works.
• Warm and smiling: Welcome back — great to have you.
• In a British accent: A proper cup of tea, if you please.
• Mid-sentence shifts: The meeting starts at nine. [whispers] But between you and me, it will run late.

Important: Bracket tags like [whispers] work for short text but fail silently with some voice+length combos (the API returns an error). Natural language directions like Say cheerfully: are more reliable across all voices.

Synthesize speech

curl -X POST "$API_URL/api/data/tts/synthesize" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, this is a sample voiceover.",
    "voice": "openai:marin",
    "instructions": "Speak conversationally, warm and relaxed.",
    "format": "mp3"
  }' | jq -r '.audio.base64' | base64 -d > voiceover.mp3

Gemini version (style direction is part of the text):

curl -X POST "$API_URL/api/data/tts/synthesize" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Say in a soft whisper: I have a secret to tell you.",
    "voice": "gemini:Aoede"
  }' | jq -r '.audio.base64' | base64 -d > whisper.wav

Multi-speaker dialogue

Gemini can synthesize a two-speaker conversation in a single call. Use this for podcasts, interviews, sketches, or any back-and-forth where the voices need to flow naturally together (Gemini blends handoffs far better than stitching separate clips).

Separate endpoint: POST /api/data/tts/synthesize_dialogue.

Hard limits:

• Exactly 2 distinct speakers per call. For 3+ speakers, split the script into 2-speaker chunks and concat the WAVs.
• Gemini voices only (gemini:<name> for every speaker). OpenAI has no native multi-speaker — use synthesize per-line and concat if you must.
• Combined text across all lines must fit the 4096-char limit.

Request:

curl -X POST "$API_URL/api/data/tts/synthesize_dialogue" \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "dialogue": [
      {"speaker": "Alice", "text": "So what do you think about the new TTS system?"},
      {"speaker": "Bob",   "text": "I was skeptical, but the inline tags actually work."},
      {"speaker": "Alice", "text": "[laughing] That must have been fun to debug."},
      {"speaker": "Bob",   "text": "[deadpan] A thrill a minute."}
    ],
    "voices": {
      "Alice": "gemini:Kore",
      "Bob":   "gemini:Charon"
    }
  }' | jq -r '.audio.base64' | base64 -d > dialogue.wav

Parameters:

Response:

{
  "success": true,
  "audio": { "base64": "...", "format": "wav", "mimeType": "audio/wav", "sizeBytes": 754604 },
  "input": {
    "provider": "gemini",
    "mode": "dialogue",
    "speakers": { "Alice": "Kore", "Bob": "Charon" },
    "lineCount": 4,
    "characterCount": 212,
    "durationSeconds": 15.72
  }
}

Tips:

• Pick voices with contrast. Kore + Charon (warm female + deep male) reads clearly; two similar voices are harder to follow.
• Let inline tags land on a specific speaker's line, not across speaker boundaries. Gemini interprets tags within the utterance they prefix.
• Keep line count reasonable per call (under ~20 lines / 30 seconds of target audio). Latency scales with output duration — a 5-minute podcast should be chunked into ~4 calls at natural scene breaks.
• For a podcast with 3+ voices, split the script by scenes where only 2 characters speak, synthesize each, concat with ffmpeg -i "concat:a.wav|b.wav" -c copy final.wav.

Parameters

Response

{
  "success": true,
  "audio": { "base64": "...", "format": "mp3", "mimeType": "audio/mpeg", "sizeBytes": 24576 },
  "input": {
    "provider": "openai" | "gemini",
    "voice": "...",
    "model": "...",
    "characterCount": 35,
    "wordCount": 7
  }
}

Decode audio.base64 with base64 -d and save to disk.

Choosing between OpenAI and Gemini

Long text (over 4096 characters)

Split at sentence boundaries, synthesize each chunk, concatenate with ffmpeg:

ffmpeg -i "concat:chunk1.mp3|chunk2.mp3|chunk3.mp3" -c copy final.mp3

Keep the same voice + style across chunks or the cuts will be audible.

Combine with video

# Replace video audio with generated voiceover
ffmpeg -i video.mp4 -i voiceover.mp3 -c:v copy -c:a aac -map 0:v:0 -map 1:a:0 output.mp4

# Mix voiceover over existing audio at 80% voice volume
ffmpeg -i video.mp4 -i voiceover.mp3 \
  -filter_complex "[1:a]volume=0.8[voice];[0:a][voice]amix=inputs=2:duration=first" \
  -c:v copy output.mp4

Gemini voices return WAV at 24kHz. Convert to MP3 before mixing into video:

ffmpeg -i whisper.wav -c:a libmp3lame -q:a 2 whisper.mp3

Delivering output

Upload finished audio to the Artifact Store so the user can access it.

Tips

• Read the voice's character column; don't just pick openai:nova by habit. The voice does more for tone than any instructions string.
• Keep instructions short — two sentences, concrete. "Speak slowly and somberly" beats a paragraph.
• For Gemini style, direct the voice like a human actor: Say cheerfully:, In a hushed whisper:, With mock seriousness:. If it would make sense on a film set, it probably works here. If Gemini rejects a style+voice combo, try a different voice or use OpenAI with instructions.
• Test one sentence first before synthesizing a long script. Adjust voice/style, then run the full text.
• Chunk long text at sentence boundaries — never mid-sentence. Mid-sentence cuts produce audible prosody jumps.