Skip to Content
Audio & Speech

Audio & Speech

Accly routes audio requests to provider models; it does not define a universal audio format. Choose a model in the Model Hub , then use the protocol and fields documented for that model. Your Accly API key authenticates the request. Provider documentation remains the authority for voice behavior, pronunciation, language support, and model-specific options.

Available routes

WorkflowAccly endpointExample modelResponse
Audio input or spoken chat responsePOST /v1/chat/completionsgpt-audio-2025-08-28OpenAI chat completion with text and/or Base64 WAV audio
Google text-to-speechPOST /v1beta/models/{model}:generateContentgemini-3.1-flash-tts-previewNative Gemini response with Base64 PCM in inlineData
MiniMax text-to-speechPOST /minimax/v1/t2a_v2speech-2.8-turboNative MiniMax JSON response
MiniMax asynchronous task creationPOST /minimax/v1/t2a_async_v2speech-2.8-turboNative MiniMax task response

The model examples above are active catalog IDs in this documentation revision. Access still depends on the current catalog, model capabilities, API key, and plan. Audio models are API-only and do not appear in the hosted chat model picker.

Active audio model IDs

The Gateway currently exposes these API-only audio models. Use the exact ID shown in your account’s Model Hub; preview and provider availability can change without changing the route contract.

ProviderActive model IDsAccly protocolOutput
OpenAIgpt-audio-2025-08-28 (alias: gpt-audio)Chat CompletionsWAV
Googlegemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, gemini-3.1-flash-tts-previewNative Gemini generateContentPCM
MiniMaxspeech-02-turbo, speech-02-hd, speech-2.6-turbo, speech-2.6-hd, speech-2.8-turbo, speech-2.8-hdNative MiniMax T2AMP3

All listed audio models are API-only. The model’s provider, protocol, and capability metadata determine which request fields are accepted; do not infer support from the model name alone.

OpenAI audio chat

GPT Audio uses the Chat Completions route. It accepts Base64-encoded PCM WAV input and can return a text answer, spoken audio, or both. Requests are buffered; audio streaming is not enabled.

OpenAI audio endpoint
POST/v1/chat/completions

Ask about a WAV recording

Replace BASE64_ENCODED_PCM16_WAV with the Base64 contents of a valid PCM 16-bit WAV file.

Audio input
curl https://api.accly.net/v1/chat/completions \
  -H "Authorization: Bearer $ACCLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-audio-2025-08-28",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Summarize this recording."},
        {"type": "input_audio", "input_audio": {
          "data": "BASE64_ENCODED_PCM16_WAV",
          "format": "wav"
        }}
      ]
    }],
    "modalities": ["text"],
    "stream": false
  }'

Generate spoken audio

The response includes text and Base64 WAV audio at choices[0].message.audio.data.

Audio output
curl https://api.accly.net/v1/chat/completions \
  -H "Authorization: Bearer $ACCLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-audio-2025-08-28",
    "messages": [{"role": "user", "content": "Read this aloud: Welcome to Accly."}],
    "modalities": ["text", "audio"],
    "audio": {"format": "wav", "voice": "alloy"},
    "stream": false
  }'

Accly currently validates WAV input up to 300 seconds and supports WAV output with the alloy voice. Keep stream false or omit it. max_completion_tokens and max_output_tokens are not accepted; max_tokens defaults to 2048 when omitted. See OpenAI’s audio guide  and GPT Audio model page  for provider behavior outside this Accly contract.

Google Gemini text-to-speech

Google TTS uses the native Gemini generateContent request shape. Accly returns the provider’s native JSON envelope; synthesized audio is Base64 data in candidates[].content.parts[].inlineData.data and the configured route returns PCM. The route is buffered and does not stream audio.

Gemini text-to-speech
curl https://api.accly.net/v1beta/models/gemini-3.1-flash-tts-preview:generateContent \
  -H "x-goog-api-key: $ACCLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Welcome to Accly."}]}],
    "generationConfig": {
      "responseModalities": ["AUDIO"],
      "speechConfig": {
        "voiceConfig": {
          "prebuiltVoiceConfig": {"voiceName": "Kore"}
        }
      }
    }
  }'

The configured voice names include Kore, Puck, Charon, Fenrir, and Aoede. The model catalog also includes gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts; availability can change. Read Google’s speech generation guide  for supported voices, prompt style, multi-speaker generation, and audio handling. Google maintains the @google/genai SDK; use it with Accly only when its endpoint and authentication configuration can target the Accly route shown above.

MiniMax text-to-speech

MiniMax uses its native T2A JSON contract, not Chat Completions. A synchronous request returns the provider’s JSON response, including generated audio on success.

MiniMax synchronous speech
curl https://api.accly.net/minimax/v1/t2a_v2 \
  -H "Authorization: Bearer $ACCLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "speech-2.8-turbo",
    "text": "Welcome to Accly.",
    "voice_setting": {
      "voice_id": "female-shaonv",
      "speed": 1,
      "vol": 1,
      "pitch": 0
    },
    "audio_setting": {
      "format": "mp3",
      "sample_rate": 32000,
      "bitrate": 128000,
      "channel": 1
    }
  }'

The configured preset voices are male-qn-qingse, male-qn-jingying, male-qn-badao, male-qn-daxuesheng, female-shaonv, female-yujie, female-chengshu, and female-tianmei. Accly accepts at most 10,000 text characters per request; stream: true, custom voice IDs, voice cloning, and timbre design are not supported.

POST /minimax/v1/t2a_async_v2 creates an asynchronous provider task and returns its native task identifiers. Accly does not currently expose a task lookup or polling route, so use the synchronous endpoint when your integration needs the generated audio in the same response.

See MiniMax’s text-to-speech guide  and T2A API reference  for provider-specific voice controls and response details.

Provider SDKs

Accly is the routing and account layer; it does not reimplement provider SDK semantics. Use the provider’s official documentation for voice names, pronunciation, language behavior, audio decoding, and model-specific parameters. The examples on this page show the wire format that Accly validates.

  • OpenAI-compatible clients can target https://api.accly.net/v1 for the GPT Audio Chat Completions contract when the client supports a custom base URL.
  • Google clients must be able to send the native generateContent payload to the Accly route and authenticate with the Accly key in x-goog-api-key; direct HTTP is the reference integration shown above.
  • MiniMax SDKs are provider-specific. Use the native T2A payload and Accly endpoint above unless the SDK supports a custom base URL and the same authentication/header contract.

Provider SDKs may expose options that Accly intentionally rejects, including streaming or unsupported model parameters. Check the selected model’s provider guide before enabling them.

Realtime and unsupported paths

The Gateway has a WebSocket GET /v1/realtime?model=... route, but the active public model catalog currently has no model enabled for Accly Realtime. Do not use the GPT Audio model above as a Realtime substitute; check the Model Hub  for availability before planning a live-session integration.

Accly does not currently register OpenAI’s POST /v1/audio/speech or POST /v1/audio/transcriptions paths. Use the provider-specific routes documented on this page instead.