​
Overview

The Talk tab in PeachDesk Voice Studio turns speech into clean text right inside your browser. Click record, speak, stop — Voice Studio transcribes what you said with a hosted speech-to-text model, optionally cleans it up with a refinement model, and gives you the finished transcript to copy or send into the rest of your workflow. Everything runs on Bibha’s GPU infrastructure — there’s nothing to install and no model files to download or manage.
Talk is the “input” half of Voice Studio’s voice loop. Cloning and text-to-speech in the Generation and Voices tabs are the “output” half. Every recording you make is also saved to the Captures tab so you can revisit the audio and transcript later.

​
The flow

1

Start recording

Open the Talk tab in Voice Studio and click Record (or use the in-app record shortcut). A recording panel appears with a live waveform and an elapsed-time counter.
2

Speak

Talk naturally — there’s nothing to wait for. The waveform and timer confirm your microphone is being captured.
3

Stop

Click Stop. The panel flips to Transcribing while the hosted speech-to-text model runs, then to Refining if auto-refine is enabled, and finally shows your transcript.
4

Use the transcript

Copy the finished text with one click, or send it into another part of your workflow. The raw audio and transcript are also paired together in the Captures tab.
Every capture appears in the Captures tab with the original audio and the transcript stored together. See Transcription for what you can do with them after the fact.

​
The recording panel

While you’re dictating, the panel walks through the states of the capture cycle and shows a live signal for each:

​
Tuning transcription

Voice Studio serves the speech-to-text models for you — there are no local files to download or GPU drivers to install. Transcription is handled by the LIPI family: LIPI by default, with LIPI-Bharat (Indian languages) and LIPI-Live (compact) offered alongside it wherever your plane serves them. Live dictation turns are decoded by the LIPI-Fast runtime automatically. What you do control is the decoding trade-off, via the beam-size setting:
  • Fast (greedy) — the lowest-latency option, and the default for casual dictation.
  • Balanced or Accurate (slow) — more decoding effort for technical content, accents, or noisy audio where transcript quality matters most.
Whatever you choose, the model runs on PeachDesk’s servers, so your results don’t depend on the device you’re signed in from.
If you frequently dictate technical terms or code identifiers, a higher beam size plus refinement (below) tends to give the cleanest results. Adding those terms to the bias vocabulary in Settings → Captures → Transcription helps too.

​
Refinement

If auto-refine is on, a hosted language model cleans up the raw transcript before you see it. The goal is to remove verbal clutter without rewriting what you actually said. What refinement typically fixes:
  • Filler words (um, uh, like used as a pause, you know)
  • Self-corrections — the model keeps the final version and drops earlier attempts (could you uh run the migration real quick, and then, yeah, check the logs → Could you run the migration, then check the logs?)
  • Basic punctuation and capitalization
  • Transcription loop artifacts — repeated tokens (six or more identical tokens in a row, case-insensitive) are stripped before the refinement model sees the transcript, so it can’t echo them back
What refinement deliberately preserves:
  • Technical terms and code identifiers (npm install, handleSubmit)
  • Legitimate repetition (no, no, no, no, no has fewer than six identical tokens, so it survives)
  • Your intent — refinement is cleanup, not rewriting
Refinement settings are snapshotted per capture, so you can re-refine the same raw transcript later with different settings without losing the original. What you tune in Settings → Talk → Refinement is behaviour, not model size: The refinement model itself is BRAHMA-30B, hosted and served by PeachDesk — it is pinned server-side, so there is no tier to pick. It is the same model that powers per-profile personality modes; see Voice Personalities.

​
When refinement is skipped

A few cases where Voice Studio leaves your transcript as-is:
  • Auto-refine is off. You get the raw speech-to-text output, which you can refine later from the Captures tab.
  • The transcript is empty or silence-only. Nothing is sent to the refinement model.
In every case the raw capture still lands in the Captures tab, so you can copy it or re-process it whenever you like.

​
Tips for clean dictation

  • Use a decent microphone and a quiet room — clean audio matters more than any setting.
  • For technical content, raise the beam size, add your terms to the bias vocabulary, and enable refinement.
  • Speak in natural phrases rather than word-by-word; the models handle punctuation and self-corrections better with context.
  • If a transcript comes out wrong, re-refine it from the Captures tab with different refinement toggles before re-recording.
Talk needs microphone access in your browser. The first time you record, your browser will ask for permission — grant it for the Voice Studio site. If recording does nothing, check your browser’s site permissions and that the correct input device is selected by your OS.

​
Next steps

Transcription

The paired audio and transcript archive every recording lands in.

Voice Personalities

The same hosted model powers per-profile compose and persona rewrite.

Generating Speech

Turn text back into natural speech with built-in and cloned voices.