Skip to content

Voice Calls and Dictation

The desktop composer provides two audio workflows: dictation creates editable text, while Call opens a live spoken conversation.

Configure voice before your first project

Open Settings → Voice from the startup dashboard to configure calls, media narration, dictation, and local speech. Call provider, model, voice and enablement save in your profile and apply across every project. A saved profile call choice takes precedence over older project call settings. If no profile choice has been saved, the panel identifies the older project choice and offers Use this voice for all projects; adopting it or changing a call preference saves the shared profile choice. Setup does not create a project or conversation. Media narration retains its own Narration settings apply to control for This project or Profile defaults. Provider keys and installed local speech packs remain shared within the app profile.

Set up a call

  1. Choose Voice settings beside the composer or open Settings → Voice. You can open settings before selecting a chat model or adding call credentials.
  2. Select Local · Qwen3 after installing local speech, or select a cloud realtime provider and add your own audio API key.
  3. Choose the Call model and Voice.
  4. Optionally choose Check model access. This checks account access without opening a paid call; it does not certify microphone, playback, or a successful conversation.
  5. Open a chat, choose Call, and allow microphone access. Calls become available automatically once the selected provider is configured; local calls also require the speech and recognition packs. Opening settings or connecting a provider does not start microphone capture or a billable session. An explicit Realtime voice calls Off setting is preserved.

Some call configuration needs an open project or an active conversation. Audio availability depends on the host, provider, credentials, and model access.

In normal chat and Orb, a connected call opens with one brief greeting after playback is ready. Greetings rotate across repeat calls and use your selected voice; they do not create a task. If you speak first, the opening is skipped. If your browser blocks playback, enable audio using the call control before expecting to hear the greeting.

Local speech and your voice library

Open Settings → Voice → Local speech and voice library. Choose Install runtime, then download the packs you need. Installation requires internet access and several gigabytes of storage. Use Verify / repair if an installed pack becomes unavailable. Speed depends on available memory and CPU/GPU capacity.

Choose Download all to queue all missing packs, or start individual downloads. Up to two packs download at once; remaining packs show their queue position and start automatically. Each pack shows its byte progress, percentage when its size is known, transfer speed and verification stage. Received network bytes can differ from pack bytes because downloads can reuse or compress data. After 30 seconds without progress, the pack shows Waiting for data with the elapsed quiet time. It remains active; cancel and retry if the wait persists. A pack becomes Ready only after verification finishes. Downloads continue when you close settings. Cancel stops active and queued downloads; completed packs are kept. Use Retry download for a failed or cancelled pack. End active local calls before starting setup, and wait for setup to finish before using local audio.

  • Preset voices uses the preset speech pack and offers nine voices.
  • My saved voices uses the reference-voice pack. Enter a name and the exact spoken words, confirm permission to use the voice, then upload or record 3–120 seconds of clear speech.
  • Generate voice sample uses the separate voice-design pack. Describe the voice and provide up to 320 characters of sample words.
  • Audition draft speaks new words in the reference voice. Listen, confirm the result, and choose Save voice. Saved voices work across your projects in this app profile.
  • Select the language, enter Preview words in that language, and use Preview selected voice. Draft auditions use these words too. Ten speech languages are offered: English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.

Call setup loads the selected synthesis and recognition models before capture starts. The first setup can take longer while these models load.

For calls, install recognition as well, select Local · Qwen3 under Call provider, and choose a compatible call model and voice. Calls become available when both packs are ready, unless you explicitly switched calling off. Local calls use turns: keep each request under twenty seconds, speak clearly, pause, and wait for the reply. Longer continuous speech is discarded and you are asked to repeat it; a partial request is not executed. Interruptions stop playback. Speed and recognition vary by hardware and microphone. Short commands and technical filenames can be misheard. For precise edits, dictate into editable text and review it before sending. Check the call transcript and use End call or repeat “Please end the call now” if a spoken hangup is missed.

For dictation, choose Local · Qwen3-ASR in Voice Input. For narrated media, select Local · Qwen3 in Media narration; recognition is required to check the spoken script. Local presets and references keep their natural expression; profiles can adjust output speed and narration sound. Script previews still require review before final narration.

After installation, local audio stays on this device and does not silently switch to a paid audio provider. Your workspace model retains its own account usage. Fully offline work also requires a local chat model and offline tools. Installation does not certify every language, microphone, voice similarity, or coding capability.

Saved voice profiles

In Local voice profiles, choose New profile or Edit profile. The editor groups voice and language, delivery speed, sound, and microphone controls into separate tabs. Save profile stores a reusable setup; Use profile applies it to calls, narration, or both as profile defaults. Existing project overrides remain authoritative. Call and narration settings also let you select a saved profile explicitly.

Speaking speed preserves pitch from 0.8× to 1.2×. Normal speed streams directly; adjusted speed buffers each short spoken segment. For a neutral outdoor sound, use Clean / open air acoustics with No added background. Room acoustics and quiet procedural rain or wind apply to previews and narration. Calls keep clean sound. Recording environment tunes local call detection; noise suppression and echo cancellation depend on device support. Longer response pauses allow more time to finish speaking.

Accent and expression come from the voice itself. In Add a voice, select English and choose an accent direction such as UK English, US English, or African, plus an optional delivery direction. These guide a new voice design; they do not retune an existing preset or reference voice. Listen to the draft and audition before saving it. Cloud controls depend on the selected model; Eleven v3 does not offer adjustable speed.

Spoken tasks and workspace context

Conversation and bounded read-only retrieval skip upfront intent classification in both Orb and normal-chat voice. Only a proposal to execute the caller’s exact completed request crosses semantic authorization before entering the persistent task runner. Existing action permissions and Orb’s scoped project checks still apply. A read-only response never acquires editing permissions.

In both Orb and normal chat, the cloud realtime model answers ordinary conversation directly. Greetings, explanations and follow-up discussion do not enter the project task queue. Requests for current information or project evidence use a separate, read-only lookup with the selected workspace model, runtime and permissions. Only an actual request to perform work can create a durable project task. Local calls still use local speech recognition and synthesis plus the selected workspace model for answers; this change does not install a new conversational model or silently switch providers. Workspace-model usage remains separate from cloud audio usage.

A lookup belongs to the call, not a single spoken turn. Accepted requests receive one brief acknowledgement of their actual state. You can keep talking or ask for a lookup’s status without cancelling it. Its verified result arrives as a named background update between exchanges and remains available for repeat/status questions during the call. Explicitly cancelling or replacing a lookup, or ending the call, cancels its pending work and approvals. With several active requests, AlphaBase asks which one you mean. Lookups stop after sixty seconds of execution time; time awaiting an approval or answer is excluded. Both lookups and tasks relay changed, observable activity. Recent useful updates suppress generic delay notices. After thirty seconds without a useful update, AlphaBase may give one fallback saying that no result is ready, without inventing a stall. A verified recovery or decision wait is explained instead. New activity takes priority over the fallback; completed results replace pending progress. Explicitly accepted project work continues independently after hangup and retains its results in the conversation.

Findings are spoken automatically when the agent returns them. Relaying an answer requires no approval or narration preview. Action permissions still apply: Full autonomy permits ordinary web lookups without a chat approval, while Ask Before Changes can require approval before using a web tool. If a task waits for permission or missing information, voice explains the wait; after you resolve it in chat, the call automatically receives the outcome. Ending the call does not stop accepted tasks.

Voice coordinates your spoken requests as one assistant. It gives brief first-person updates, such as “I’m checking current information,” during longer tasks, then explains the result as its own work. Updates reflect actual task activity; quick answers skip progress announcements. Findings retain their sources and limitations. When permission is needed, it says “I need your permission to continue.” You can ask how the voice and workspace models divide the work.

The voice session receives workspace context and bounded conversation updates. It does not automatically load every file or every past conversation. Name the relevant task or file when needed. Switching agents does not change the panel’s neutral Voice call title.

Use End call or explicitly say “end the call.” Ending audio can leave an already accepted task running; stop that task separately when necessary. Expand Transcript and scroll to later speech. A saved recording can be played or downloaded when capture succeeds.

Choose Balanced · 0.6B or Quality · 1.7B recognition in local speech settings or a saved voice profile. Download the selected pack first. Quality uses more memory and may take longer; your selected chat model continues to perform tasks. Existing profiles keep Balanced by default.

Local calls speak complete answer sentences as soon as the workspace model explicitly identifies its response. Other models wait for a confirmed answer boundary. Explicit hangup phrases cover the ten local voice languages. If a short command is unclear, the call may ask whether you meant to end it; answer yes or no within thirty seconds, or use End call. This does not add approval steps to ordinary answers. Actual microphone, accent and noise performance still depend on your device and require testing.

Recording retention

Saved recordings remain on your device. Deleting a conversation removes its history and playback references; cached audio and files you downloaded may remain. Provider-side retention follows the provider’s policy. Conversation deletion does not guarantee erasure of every audio copy.

Dictate a message

In Settings → Voice → Voice Input, choose a transcription provider, the expected language, and a microphone. Offered choices are Auto (prefer OpenAI, then Groq), OpenAI Transcribe, Groq Whisper, and Local · Qwen3-ASR. Install the local recognition pack or supply API access for your chosen cloud service.

Use the microphone control, record, then review the transcript before sending it. Dictation does not execute the task until the text is sent. It uses its own provider rather than the active chat model. Local audio stays on this device. Cloud audio is sent to that transcription provider and is not attached to the chat as an audio file.

Access, cost, and troubleshooting

ChatGPT and GitHub Copilot subscription logins do not provide audio API keys or included AlphaBase voice minutes. Cloud calls and dictation use the audio provider’s access and charges. Local speech has no hosted audio API charge; it uses device compute and storage.

If setup fails, read the error, check the selected audio provider and credentials, and run the available model-access check. For recognition problems, verify the input device and language and reduce background noise. If a spoken task appears stuck, inspect its Activity and pending approvals; ending the call is not a substitute for stopping that task.