← Voice Lab · Explore

Voice Lab quick start

Recorded audio

  1. Set Listen from to Audio file and choose your recording. Select a passage up to 60 seconds.
  2. Select the spoken language or Auto-detect, then English, Spanish, or French output.
  3. Click Enable translation online once, then Translate from playhead to process up to 60 seconds from the player position. Use play/pause, seek and volume on the player. Stop / rewind resets playback. Seeking cancels in-progress file translation to avoid stale text; enable cached models again afterward.
  4. Replay passages, compare machine text with the sound, and save corrections in the notebook. Export a backup; browser storage can be cleared.

Connected radio or microphone

  1. Connect the radio USB audio interface and choose Connected radio / microphone under Listen from.
  2. Click Find audio inputs and allow audio permission to list device names. Select your radio USB input. Selecting a source alone does not start recording.
  3. Select languages, then click Listen to input to hear incoming audio. Monitor volume controls speakers without changing the translation signal. Enable translation to process live passages. Stop listening releases the input and cancels pending translation.
  4. Record a clip instead is optional: Record input, Stop recording, then Translate from playhead. Recording stops at 60 seconds. Use headphones to avoid speaker-to-microphone feedback.
  5. This captures receive audio only. It does not tune the radio or control transmission. A disconnected selected device reports an error rather than silently using your microphone.

Live Internet DMR

  1. Choose Internet DMR, search by country/region, talkgroup name or number, choose a talkgroup from the dropdown, and click Play DMR. A green dot plus Active means live transmission reported by BrandMeister; gray means no current activity or unavailable activity data. Green clears when the live feed disconnects or becomes stale. Country/region is inferred from the talkgroup numbering plan, not the caller’s location. Listening works before downloading translation models.
  2. Player volume controls listening. The level meter measures actual received audio. Waiting means connected with no current transmission; it is not a guarantee of speech on that talkgroup.
  3. Select languages before playing, then Enable translation. The same decoded audio feeds local translation in bounded passages; short calls are processed when the transmission ends. Models may lag or skip passages on slower devices.
  4. Stop DMR closes the receiver and cancels pending translation. Choosing a new talkgroup stops the old receiver. Click Play DMR to listen to the new selection.
  5. If the native receiver is unavailable, Hoseline fallback opens its player. Fallback tab capture requires a supported desktop browser and explicit sharing with audio. Native DMR needs no tab sharing or microphone permission.

Offline preparation

Under Advanced & offline tools, open Offline settings and run Check offline readiness for each output language. Then disconnect and reopen the page and process a file to verify your own browser. Hosted sign-in can require internet; cached models alone do not guarantee offline reopening. Live Hoseline cannot work without internet. No paid transcription API is used.

Built-in self-test

Run sample self-test loads a bundled 11-second English excerpt, transcribes it and generates Spanish output in a separate local worker. It may download Spanish models on first use. Exact normalized English word comparison can fail even when both engines run; Spanish reference cues are a limited smoke check. Export self-test log includes stage results, timing and edit counts, not communication content. Cancel terminates the worker. Normal processing needs model preparation again afterward.

Run cache-only test reads the sample from browser Cache Storage and restricts model fetches to cached responses after the test worker starts. Missing cached files fail instead of falling back to network. Worker/module startup and the whole site’s authentication/navigation are outside that check; it does not prove disconnected site reopening. HTTP previews without secure browser caching report unavailable.

Diagnostics and reports

Diagnostics measures selected audio level (RMS in dBFS, relative to digital full scale), model state, last processing duration, pending passages and skipped audio. A nonzero level does not prove intelligible speech. Processing slower than the audio duration warns that live decoding may fall behind. Export test report downloads settings and measurements with categorized errors, excluding audio, transcript, translation, filenames and notebook content. Reports stay local until you choose to share the exported file. These are operational measurements, not accuracy or confidence scores.

Running text

The source-language transcript and translation accumulate as paired, timestamped passages. Stop and source changes retain the session text. Each window scrolls independently. Uncheck Follow latest to read older text. Save text exports the accumulated session; Clear session removes it. Reloading or closing the page does not preserve unsaved session text.

Live audio queues while translation processes. If local processing cannot keep up, both windows show a skipped-audio marker. Repetitive model output is flagged and withheld instead of being presented as reliable speech. A warning is not a corrected translation. Select Spanish explicitly when the spoken language is known; country labels do not set the speech language.

Accuracy and review

99 input languages are offered by the model; accuracy varies. Spanish and French use an English intermediate translation, which can compound errors. Noise, overlapping voices, call signs, numbers and short clips need particular care. Review flags are qualitative prompts, not calibrated confidence scores. Vocabulary entries only help review exact text matches; they do not train the engine.

Privacy and responsible use