ドキュメント / 使い方
ドキュメント · Set up your services
OPENLESS / HANDBOOK

Set up your services

Start with Traditional: ASR transcribes your voice, then an optional LLM edits the words. Alternatively, go straight to Multimodal or the DeepSeek suite. You only need one pipeline.

Configure your ASR

In Settings → AI services and models, select Traditional. ASR turns speech into a transcript. Choose one provider below, save and verify the channel, enable it, and put it first among enabled ASR channels. For a plain transcript, you can skip the LLM.

Volcengine

  1. Activate speech recognition

    Sign in to Volcengine. In Ark, go to System management → Service activation → Speech models and enable Doubao streaming speech recognition. You can also use the Doubao Speech console. Check that the service is active.

  2. Get your credentials

    In the new console, create an API key under API management in the same project. In the older console, select the model and application in the API service center and copy the APP ID and Access Token from the authentication details. The Secret Key shown there is not the Access Token.

  3. Connect OpenLess

    Open Settings → AI services and models → Speech recognition → Add channel. Choose Volcengine and the standard service. Select API Key or APP ID + Access Token to match your console, fill in the corresponding fields, save, and verify.

For standard streaming ASR, you can leave Resource ID blank. If you need to enter it explicitly and activated the matching duration-based service, use volc.seedasr.sauc.duration. If access is denied, check your service entitlement. For Agent Plan, select that plan and use its dedicated key.

Official guides: Console guide · OpenLess setup notes

Alibaba Cloud Model Studio

  1. Choose a region

    Sign in to Model Studio and activate the service. Select the region and workspace that host your model. For the default mainland China endpoint in OpenLess, start with the Beijing region.

  2. Create an API key

    Open the API key page for your region (Beijing region; for other regions, see the official guide below). Create a key for the appropriate workspace and copy it. Use a Model Studio (DashScope) API key, not an Alibaba Cloud account AccessKey.

  3. Choose a speech model

    In OpenLess, add a Model Studio channel under Speech recognition and paste the key. Start with the default fun-asr-realtime model, then save and verify.

The key, endpoint, and model must belong to the same region. If verification fails, check the region and your account’s access to the selected model.

Official guides: Official key guide

iFlytek

  1. Choose an application

    Sign in to the iFlytek console. Under My applications, open an existing app or create one.

  2. Activate real-time transcription

    Enable standard real-time transcription (RTASR) for this app. Open that service’s authentication details and copy its APPID and APIKey. Credentials for a different speech product won’t work here.

  3. Connect the same app

    Add an iFlytek channel under Speech recognition in OpenLess. Enter the AppID and API Key from the same app, save, and verify.

If authentication fails, make sure RTASR is active, both credentials belong to the same app, and your device clock is correct.

Official guides: RTASR documentation

Configure your LLM

An LLM edits the text returned by ASR: it can remove fillers, adjust tone, or format a draft. Mix providers if you like, such as Volcengine for speech and DeepSeek for editing. Add a separate language model channel; a text model ID does not belong in the ASR field.

DeepSeek

  1. Create an API key

    Sign in to the DeepSeek Platform and open API Keys. Create a key, give it a recognizable name such as OpenLess, and copy it.

  2. To edit text: add a language model

    In Settings → AI services and models → Language models, add DeepSeek. Paste the API key, select a model available to your account, save, and verify. In the traditional pipeline, transcription uses a separate ASR channel.

  3. Test the edit

    Enable and select the language model channel and an editing style. Dictate “Um, the draft is due Friday—no, Thursday.” Check the corrected day and the tone, not just whether text appeared.

This configures text editing only. For DeepSeek speech recognition, Q&A, and read-aloud, see the DeepSeek suite.

Official guides: First API call

Volcengine Ark

  1. Prepare a text model

    Enable a text model in Ark, create an Ark API key, and copy the model or inference endpoint ID shown in the console. The APP ID and Access Token from the speech service are not substitutes for this key.

  2. Add an editing channel

    Under Settings → AI services and models → Language models (text editing), add Ark. Choose the service you activated, enter its key and model ID, keep its preset endpoint, then save, verify, and enable the channel.

Standard and subscription services use different keys and endpoints. If the model is not found, check the service type and endpoint ID.

Official guides: OpenLess configuration

Multimodal dictation

A multimodal model handles more than one kind of information. It might transcribe a recording, describe an image, or summarize a video. It need not support all three. For dictation, check two things on the exact model card: audio input and text output. A vision-only model can be multimodal without being able to hear you.

Traditional mode sends the transcript from ASR to a separate LLM. Multimodal mode sends a recording and instructions to one model and gets text back. This OpenLess workflow uses the microphone; it is not an image or video upload feature.

  1. In Settings → AI services and models → Recognition pipeline, choose Multimodal. Its settings are separate from traditional channels. Confirm the key and model under Multimodal model.
  2. Choose a provider below. Set the provider, model ID, API key, and base URL. Check that your account can access the example model.
  3. Save and verify. In a blank note, hold the recording shortcut and say “Tomorrow at two—sorry, three.” Release, read the result, and check the corrected time. Verification for general models may test only text connectivity, not audio input.

If text works but audio is rejected, check the specific model and endpoint for audio support, then check access and quota. Switch back to Traditional to restore your previous ASR and LLM settings.

Google Gemini

  1. Choose a model

    For example, gemini-3.5-flash-lite. Gemini Flash models support several kinds of input. Check the exact model card for audio input and text output.

  2. Fill in the Gemini preset

    Create a project API key in Google AI Studio. In Multimodal model, choose Google Gemini, enter the model ID and key, and keep the default base URL https://generativelanguage.googleapis.com/v1beta. Save and verify.

If your release offers dedicated Gemini Transcribe models, they use Smart transcription, not general Q&A or translation.

Official guides: Audio input guide

OpenAI Audio

  1. Choose an audio model

    For example, gpt-audio accepts a recording and can return text. Not every GPT model accepts audio on the same API; image support alone is not enough.

  2. Fill in the OpenAI preset

    Create a project API key on the OpenAI API platform. Choose OpenAI (audio), enter the model ID and project key, and use https://api.openai.com/v1 as the base URL. Save and verify, then test a recording.

Your API project needs model access and available quota. Use an API key, not a ChatGPT login password.

Official guides: GPT-Audio model card

Alibaba Cloud Qwen-Omni

  1. Choose an Omni model

    For example, qwen3-omni-flash. Qwen-Omni can understand text, audio, images, and video. For dictation, choose an audio-capable model available in your region.

  2. Fill in the Model Studio Omni preset

    Choose the model’s region and workspace in Model Studio, then create an API key (Beijing region). Choose Model Studio Omni in OpenLess, then enter the key and model ID. Use the compatible-mode base URL for that region; the mainland China preset is https://dashscope.aliyuncs.com/compatible-mode/v1. Save, verify, and record a sample.

Do not paste the traditional ASR WebSocket endpoint here. Omni uses the model service’s compatible API. The model, key, and endpoint must match the region.

Official guides: Model Studio API keys · Qwen-Omni inputs

DeepSeek suite

Choose DeepSeek suite to use DeepSeek for speech recognition, Q&A, native web search, and read-aloud. It combines separate DeepSeek services. This is distinct from using DeepSeek just to edit ASR text or using one audio-capable model in Multimodal mode.

  1. Create an official key on the DeepSeek Platform and check your API balance.
  2. In Settings → AI services and models → Recognition pipeline, choose DeepSeek suite. On first setup the key may be copied from an enabled official DeepSeek text channel. If the field is empty, paste your key. Choose a voice and save. Subsequent key changes must be saved in the suite itself.
  3. Dictate a sentence in a note and release the shortcut. Upload starts after recording stops, so no live words are expected. You can also ask a question and use Read aloud on the finished answer.

If your release does not show the suite, use Traditional ASR with a DeepSeek LLM instead. Speech and web use are billed by the provider.

Official guides: OpenLess suite notes · DeepSeek speech API

Try a short recording

  1. Save and verify the channel. If you have several, make sure the configured channel or suite is selected.
  2. Open a blank note, hold the recording shortcut, say “Meeting tomorrow at three,” and release. Check that the words appear in the note.

No text at all? Check microphone access, the ASR channel, and your provider quota. If transcription works but editing does not, check the language model channel. Enter keys only in the app or provider console; keep them out of chat, issue reports, and screenshots.