Interpretation workbench

Unified interpreting speech settings

Under Interpreting speech, choose a local native model, cloning model service, speech service/API or compatibility settings. Selecting a route enables only that route while preserving other models, references and credential drafts. CosyVoice3/Fish have local and service options here. Local execution requires a complete matching native package and runtime; actual native generation is not yet verified, and the official original checkpoint folder cannot be used directly.

Continuous interpretation and Mini speech output do not add audio watermarks, avoiding watermark processing and buffering. Ordinary cloned recordings saved from Voice Tools may still contain provenance watermarks. Playback latency also depends on the models, translation service and audio devices.

If an older configuration enables multiple routes, a warning appears. Explicitly choose one route before starting the task. Route switching is locked during preparation, running, stopping and finalization.

1. Save local model paths once in Setup guide. Give each size or quantization a clear name and variant; save a VoxCPM runtime with its model when needed.
2. In the workbench choose Run using → Local native model, then choose a saved Continuous cloned speech model. Do not enter its path again. Manage local models opens Setup to add or change registrations.
3. The task uses the model currently selected in the workbench; the main window can have a different saved model. After choosing another registered entry, check its displayed name and variant, then try a short passage to check recognition and speech output. Choose a saved Voice library reference and the compute device for this task. Check the model and reference, then audition a short target-language passage before continuous interpretation.

1234

1. 1Choose Local native model.
2. 2Choose a registered model; its name and variant distinguish versions.
3. 3Use Manage local models to add a model or edit its path in Setup.
4. 4Choose a saved reference recording; the voice can vary by task.

123

1. 1Choose a cloning model service.
2. 2Choose CosyVoice3 or Fish.
3. 3Use a reference you have auditioned.

Check three stages first

1. Recognition: confirm input, STT backend, model directory and source language; verify source text with a short passage.
2. Translation: save the endpoint, model and authentication in Setup guide and test a short text. In the workbench, choose the saved Translation service and target language.
3. Synthesis: choose a TTS backend or explicitly enable native/service voice settings; audition a short target-language passage first.

12

1. 1Select NVIDIA GPU, or keep CPU.
2. 2Leave the CUDA field empty for automatic discovery, or choose a file.

Task languages and preceding source context

Requires the Windows r512 or later installer preview. Earlier previews do not include these language and context fixes. These settings shape translation requests; review the original text, translation and generated speech yourself.

1. Before starting, check the task's Source language and Target language in Interpretation Workbench. For Chinese speech and English translated speech, select “Chinese → English.” A Chinese or English interface does not replace these choices. Check recognition, translation and speech separately before starting the run.
2. If earlier wording would help resolve references or terms, expand Terminology and translation context in translation settings. Set Recent source context budget (0 disables) to a suitable value above 0, save, then begin a new task. The default 0 sends no preceding text; individual segments can still be translated with it left at 0.
3. With context enabled, segmented speech translation and segment-by-segment translation of saved audio reuse original text from earlier successfully translated segments in the same task. For example, “This is Boundless Flow” can provide context when the next segment says “It can record meetings.”
4. Context helps with the current segment. It does not include preceding translations, is not translated again and is not shared across tasks. A new task begins with its own context. If the budget cannot fit all preceding text, only recent source text that fits is included; the current source is not truncated to make room.

Enabling context sends that preceding text to your selected translation service. Check glossary and request (no model call) in the main recording settings uses that page's current recognition settings and primary target, with no session history. It is not a replay of an interpretation task's full history. After changing a language, backend, target or model, run the check again and review the task choices in Interpretation Workbench. See source, target and interface languages.

Continuous interpretation

Open Interpretation Workbench from the header and inspect source and target languages, additional subtitle languages and output. Start continuous operation and observe preparation. Components supporting preloading prepare the voice model first; otherwise the UI indicates first-utterance loading. Stop can cancel preparation.

The workbench plays generated speech automatically. With Rust streaming playback or model-service Play audio as it arrives, the default output device is prepared before recording and reused throughout the run. Stop releases it and can also cancel preparation. A default-device change or interruption stops the run and displays an error. Check the device and restart interpretation.

During operation, inspect segment text, translation, synthesis, playback and backlog. Use available catch-up actions to handle backlog; skipping playback does not mean deleting the transcript or translation. Audition a short passage first and avoid feeding playback back into the recording input.

12

1. 1Start is primary. Stop stays here throughout the session.
2. 2Read current status. Closing the panel keeps the session running.

12

1. 1Recording begins only after the output device is ready.
2. 2Use Stop to cancel during preparation.

Cloned voices and output

Native cloning requires a prepared and auditioned local model, reference audio and its matching transcript. Service cloning uses separate service configuration. A reference does not guarantee cross-language voice consistency; test the target language. Do not switch models or move files during a running task.

Continuous STS (experimental) and Stop Live STS stay at the top of the panel while you scroll through settings. Closing the workbench hides it while processing and playback can continue. Select Stop Live STS to end the task, and confirm recording finalization, synthesis cancellation and playback separately.

Streaming speech and latency

First save and audition the speech endpoint, credentials, model and voice under Setup guide → API Config. In the workbench choose Speech service/API and a saved provider without entering the connection again. Voice, emotion, speed and playback can still be adjusted for the task. Follow the displayed requirements for complete or incremental text.

Model loading and warm-up belong to preparation. Before continuous interpretation, wait for components that support preloading to finish preparing. Other components may load on the first utterance, which should be assessed separately. After preparation, speech-to-speech latency still depends on segmentation, recognition, translation, first audio, queues and devices. Try a short passage before a longer session.

During operation, check Source start → first audio and Source end → first audio. The former includes the spoken passage; the latter helps assess the response after its end. Recognition after source end includes segmentation wait; Translation observed includes result delivery to the interface. Last enqueue-to-first-audio covers synthesis, queueing and other work after enqueueing, excluding earlier recognition and translation. Total request time cannot replace these measurements.

Native playback prefers the driver timestamp for its first audio payload. Web Audio and media playback use scheduling or playback events; older native components use software notifications. These are software estimates, not measurements of sound from the speaker. Clock alignment reports clock uncertainty only; devices add their own uncertainty. Missing, expired or mismatched timestamps show — rather than zero. Autoplay disabled or failed playback produces no new measurement.

The screenshot shows a waiting state with no latency measurements yet. Services supporting streamed PCM can begin playback before the response ends; newer components deliver audio directly when Rust output is selected. Models returning only a complete WAV still require full synthesis. A streaming playback setting does not make such a model incremental.

Streamed audio resumes delivery directly after a completed background wait. Stop can interrupt waiting; older components retain their waiting safeguards. Actual response time depends on the model and device. Check runtime measurements and audition the result.

Run statistics show sample counts, p50, p95 and maximum latency from source start and source end to first audio. Percentiles use valid timing in the latest 256 settled speech tasks; completed, failed, stopped, skipped and unavailable source timing are counts for the whole run. Valid first audio remains when a task later fails or stops. Missing or unplayed timing is not replaced with zero. Recognition or translation failures before enqueue and frontend admission rejection are outside these statistics. Stop retains the report; the first queued task of a new run clears it. Keep first utterances, reconnects and slow segments when comparing the same models, languages and devices. These values are software timings.

123

1. 1Choose a speech service or API.
2. 2Choose the provider matching your service.
3. 3Enter a voice available through this service.

123

1. 1Read source-start and source-end latency separately.
2. 2Check recognition after source end and observed translation time.
3. 3Missing clock alignment produces no estimate.

Play model-service audio as it arrives

The CosyVoice service must return mono 24kHz PCM16. Fish must return mono PCM16 WAV at 16–96kHz with the official 44-byte header. First audition a short text to check the service and reference, then enable Play audio as it arrives. First-audio latency can improve only if the server yields audio early. A server returning audio after full generation will still wait for the complete clip.

Each finalized translation is sent once. Incoming audio can play before the response ends; the completed audio is saved afterward. Fish streaming audio is converted to 24kHz for playback and saved at that converted rate. Disable the option to use complete WAV at its original sample rate. The original reference recording is unchanged. Interrupted or invalid audio is not saved as a complete clip, and already played sound cannot be recalled. This option does not accept incremental text.

1. First save and audition the CosyVoice or Fish server root and key in Setup guide. In the workbench choose Cloning model service and the saved service, then select the reference recording and transcript for this task.
2. Enable Play audio as it arrives, read the requirements, then start continuous interpretation. Disable it to retain complete-clip generation before playback.
3. Check first-audio latency, continuous playback and voice quality on your devices. Stop interpretation to pause generated speech; replay the completed audio from the results.

123

1. 1Select the running cloning model service.
2. 2Enable audio-as-arrived playback; complete-clip playback remains available.
3. 3Latency improves only when the server yields audio early.

123

1. 1Select the running cloning model service.
2. 2Enable audio-as-arrived playback; complete-clip playback remains available.
3. 3Latency improves only when the server yields audio early.

123

1. 1Completed, failed and stopped are run totals; missing source timing is not zero.
2. 2Valid source-start timing in the latest 256 tasks; slow samples are retained.
3. 3Source-end sample count, p50, p95 and maximum are separate.

FireRed generation-time playback · Experimental

Register and save the original FireRed root in Setup guide, then audition FireRed Instruct briefly in Voice Tools. Under Interpreting speech choose Local native model, that saved FireRedTTS3 entry and CPU. The current Windows package does not provide FireRed CUDA components. The original root must contain fireredtts3_instruct, redae and text_tokenizer; no separate server or Index conversion is needed. Keep the full app folder with matching runtime components. Prepare native model does not supply missing runtime components. After checking the model and reference, supporting components offer Prepare model and reference. Wait for preparation, then audition a short sentence. Recheck and audition after changing the reference or transcript before restarting interpretation.

Use a 1–10.24 second reference, preferably 3–5 seconds of clear single-speaker speech, with its actual transcript. Enable Play while generating · Experimental. The workbench plays generated speech automatically. Start continuous interpretation and wait for model, reference and output preparation before speaking. Stop can cancel preparation.

Each finalized full translation starts generation. PCM goes directly to native playback during generation without waiting for a complete WAV. It does not accept arbitrary appended text, and early-sentence synthesis is hidden while selected. Preparation does not establish realtime runtime performance. Observe backlog and first audio during continuous speech and listen on your devices. Sustained realtime, physical devices and full voice quality remain unaccepted.

Stop cancels preparation or generation and releases output. Closing the panel does not stop interpretation. To return to complete-WAV playback, stop first and disable generation-time playback. If the component supports only complete WAV, use the full app folder with updated FireRed components.

The older screenshot illustrates control locations. Its CUDA setting is from an earlier setup; select CPU for FireRed in the current Windows package. For NVIDIA GPU use a supported Index-TTS2, Qwen3-TTS or VoxCPM2 route after checking its components and device.

1234

1. 1Use a 1–10.24 second reference and transcript; preferably 3–5 seconds.
2. 2This control shows CUDA in the older screenshot; select CPU in the current Windows package.
3. 3Choose FireRedTTS3.
4. 4Select the registered FireRed model; its original folder is saved in Setup guide.

12

1. 1Enable Play while generating · Experimental.
2. 2Automatic speech is also required; sustained realtime remains unverified.

Early native sentence synthesis · Experimental

Under Interpreting speech, select Local native model, check the model and reference recording, then enable Synthesize translated phrases early · Experimental. It is off by default and can be changed after interpretation stops.

After a source speech segment is finalized, checked translation phrases can enter synthesis early, with the next phrase prepared during playback. Chinese, English and other targets use checked phrase boundaries; legacy text still waits for clear sentence boundaries. Full text and replay stay with the original segment. Recognition drafts are not spoken.

Checked phrases are generated in order, even when translation finishes before synthesis starts. A remaining tail without checked boundaries can be combined into one request after final translation when it fits within 512 characters; longer tails still use sentence boundaries. Full text and replay are preserved.

The model still generates a complete WAV for each request. Splitting may provide playable audio sooner, increase total synthesis time or change pauses. Test your target language and reference voice, checking first audio, continuous playback and voice quality. If the result is unsuitable, stop interpretation and turn the option off.

12

1. 1Enable early sentence synthesis under Local native model; it is off by default.
2. 2Final text is combined within size limits. Each request still produces a complete WAV. Listen for pauses and voice quality.

After stopping, wait for records to finish saving

Select Stop Live STS to stop translated speech immediately. Recording finalization and record saving may continue afterward. If Saving interpretation records appears, wait; stopped playback does not confirm a completed save.

If records remain unsaved, check free space and write access at the save location, then select Retry saving. The app prevents starting another task, restarting the app or restoring a different draft while this recording still has unsaved work. Finish saving this run first.

To leave the window, choose Close interface · Keep working in the exit dialog. This hides the interface and keeps the current app process running. Keep the computer on; do not quit the app, restart it or shut down before saving completes. Closing the interpretation panel only hides that panel and may leave an active run working.

To abandon the unsaved part, explicitly choose and confirm discarding this draft. Hiding the interface does not discard it. Unsaved work is retained only in this running app; do not rely on recovery after a restart.

12

1. 1Unsaved interpretation records remain in this running app.
2. 2Retry saving after storage becomes available.

Save CosyVoice once, then choose an account voice

These CosyVoice connection steps require the Windows r512 or later installer preview.

1. Open Setup guide → API Config and enable Alibaba Cloud · CosyVoice. This card is for text-to-speech only. Enter the complete wss:// endpoint for your own workspace and region, ending in /api-ws/v1/inference. Replace any WorkspaceId placeholder, then enter the matching API Key and TTS model. Complete the guide to save.
2. Return to Interpreting speech → Speech service / API and explicitly select Setup · Alibaba Cloud · CosyVoice · Incremental text. Check the model and host. The app does not guess a region, copy an Aliyun text or Qwen Audio 3.1 connection, or start interpretation automatically.
3. Enter a voice ID that matches the model, or choose an account voice below. The voice starts empty when you first select this source. The speed range is 0.5–2. Use supported instructions for emotion; there is no separate emotion parameter. Custom connections and other sources keep their own usage drafts.

Get the complete endpoint from your Alibaba Cloud account. See the official WebSocket guide for its format and speech model reference for model, voice and instruction support. Account access, region, model and voice must match.

Find an existing voice before creating one

1. Click Clone / design / select voices in STS. The panel opens on My voices and shows the saved host and model without duplicate connection fields. Click Find cloud voices to query the account. Opening the panel does not query, upload or create anything.
2. For a new voice, select Design from text or Clone a recording, fill in the requested fields, then click the creation button. cosyvoice-v2 supports account queries, existing voices and reference cloning; text design is disabled with an explanation. cosyvoice-v3.5-plus and cosyvoice-v3.5-flash require a created cloned or designed voice and have no preset voices.
3. Select Check status, then Generate preview. Listen to the complete preview before selecting Use for interpretation. Returning to STS changes only the voice ID. The saved source, instructions, speed and playback choice stay selected, and custom-connection drafts remain separate.

Instructions allow at most 100 weighted characters: Han ideographs count twice and all other characters once. For example, 50 Han ideographs reach the limit. A voice must match the target model used to create it; check the voice again after changing models. If the service rejects an instruction, its draft remains available to correct.

Queries, creation and previews contact the selected service; creation and synthesis may incur charges. After Cancel or a change to the saved connection, wait for the original request to finish before the controls unlock or the new connection applies. If creation is unconfirmed, query the original account before creating again. The app does not automatically retry creation or switch away from a disabled or removed source.

Continuous mode can supply translated text progressively and receive audio as it arrives. Mini still waits for you to stop the clip, then translates and generates complete audio before playback. Try short text with your own account, target language and output device before a meeting.

123

1. 1Use the connection saved in Setup. Manage connection opens the guide.
2. 2Find cloud voices runs only when clicked. My voices is the initial tab.
3. 3Choose a voice and generate a preview. Use for interpretation requires the complete preview first.

123

1. 1The saved CosyVoice source stays selected.
2. 2Only the voice ID changes to the auditioned account voice.
3. 3Return to account voices here without re-entering the connection.

Save MiniMax once, then choose its speech mode

1. In Setup guide → API configuration, enable the MiniMax card, enter the service address and API key, and choose its TTS model. Finish the guide to save. Opening the card alone does not start interpretation or create a voice.
2. Return to STS → Speech settings → Speech service / API. Explicitly choose Setup · MiniMax · Complete text or Setup · MiniMax · Incremental text. Check the displayed model and host. A saved full endpoint may fit only one mode; fix the address in Setup if the other mode is unavailable.
3. Choose or enter an account voice ID before starting. A new MiniMax source starts without a selected voice. Both modes share the saved address, key and model, but retain separate voice, emotion and speed choices. Switching modes does not overwrite the other mode’s choices or your older custom connections.

The protocol distinction follows MiniMax’s HTTP speech API and bidirectional WebSocket API. Mini still waits for you to stop recording, then generates a complete audio file before playback, whichever source you choose.

Find, clone or design an account voice

1. Click Manage voices under the STS voice settings. The MiniMax voice panel automatically uses the selected saved connection and shows its host and model; it does not ask for the URL, key or model again.
2. Click Find voices to query the account. For a new voice, open Clone or Design, fill in the requested recording or description, then click the explicit creation button. Opening the panel does not query, upload, create or synthesize anything.
3. Click Audition for a voice. Listen to the complete result, then click Use for interpretation. The current source keeps its protocol, emotion and speed; only its voice changes. Return to STS and check the selected source and voice ID.

Voice creation and synthesis may incur charges; a reference is uploaded only when you click the clone operation. Cancel waits for the current operation to finish before releasing its controls. If creation is unconfirmed, keep the voice ID, restore local records or audition that ID using the original connection, and avoid submitting creation again. A disabled, removed or changed saved connection requires attention in Setup; it does not fall back to a custom connection.

Try short text with your own account, target language and output device before a meeting. Real cloud sound quality and latency have not been accepted by the controlled examples below.

123

1. 1The saved connection supplies the host and model. Manage it in Setup.
2. 2Click Find voices to query the account; opening this page does not query it.
3. 3Choose a voice and click Audition. This inventory is a controlled example.

123

1. 1The original Complete text or Incremental text source stays selected.
2. 2Only the auditioned voice changes; this voice ID is an example.
3. 3Use Manage voices to query, clone, design or audition without re-entering credentials.

Use a speech API in Mini

The same choice under Interpreting speech → Speech service / API applies to both continuous segment speech and Mini. This route needs no local TTS model folder or reference recording. Use a preset or previously created voice ID available to your account.

123

1. 1Choose the saved service connection. Endpoints and keys are managed in Setup.
2. 2Enter a voice ID available to your account; coral is the example shown here.
3. 3Mini records a clip, then translates and generates complete audio after Stop. Wait for cancellation to finish before recording again.

1. Save the speech connection and key in the Setup guide, then audition a short passage.
2. Return to Interpreting speech, choose Speech service / API and the saved connection, then select a voice. Adjust expression controls only when the service supports them. You do not need to enter the endpoint or key again.
3. Open Floating Mini and select Start. Speak after Recording appears and the input meter responds, then click again or press RightAlt to end the clip.
4. Wait for translation and speech generation. Mini generates a complete WAV before playing it. Even when a service supports streaming, you must first stop recording this clip.
5. To cancel, select Stop in the workbench or Exit Mini. Wait for network requests and file saving to finish stopping before recording again. Wait for the task to be confirmed ended before changing its settings.

Available adapters include OpenAI PCM, MiniMax SSE / WebSocket, CosyVoice, Doubao SSE / WebSocket, ElevenLabs and the VibeVoice service. Availability depends on the saved connection and account permissions. Use Qwen3.1 TTS Next in Voice Studio; that adapter is not available in Mini here.

If an older configuration enables several speech routes, explicitly choose one first. Invalid connection, voice or expression settings produce an error rather than silently switching to a local model. Translated text is sent to the selected speech service. If recognition also uses an API, recorded audio is sent to that recognition service. Mini and continuous interpretation add no audio watermark. Test voice quality and waiting time with your own account, languages and devices.

123

1. 1Explicitly choose Setup · OpenAI TTS and check the displayed model and host.
2. 2Choose a voice ID for this connection. The default shown here is alloy.
3. 3The same connection works in Mini after recording stops; Mini waits for complete audio before playback.

Mini locks service, voice, language and configuration choices while awaiting start confirmation, recording, finalizing the clip, translating, generating complete audio and playing it. After selecting Stop, wait for confirmation that the task has ended. If cancellation is unconfirmed or fails, the choices remain locked; select Stop again to retry. Change settings or record another clip only after the task is confirmed ended. You can still stop recording, generation or playback, collapse cards and view their content.

123

1. 1Run using remains locked while the Mini task is active.
2. 2Stop the active task before changing settings. Wait for end confirmation; retry Stop if cancellation is unconfirmed or fails.
3. 3The selected voice ID is retained and cannot be edited until the task is confirmed ended.

Choose once, use in each workspace

Register each local model in Setup with a recognizable name and its size or quantization in the variant field. Save separate profiles for different variants. VoxCPM runtime folders are saved with the model.

Workspaces show the model’s purpose, name and saved variant. Older entries named after a backend, such as ONNX, show their folder name to help tell them apart. A displayed name does not verify the files or their capabilities; use Check model before a short trial.

In the translation studio, speech recognition converts input to text. For local speech routes, segment speech voices a completed Mini recording, while continuous cloned speech uses the selected model and reference. For Speech service / API, Mini and continuous segment speech use the same saved connection and voice. Select the route for your task. Manage models opens Setup; there is no need to enter the path again.

For supported continuous parameters, turn the dial or focus it and use arrow keys. The number field remains available for precise entry. During operation, unavailable controls are dimmed and cannot change the running model.