Speech & voices

Choose an operating mode

Open Voice Tools → Model Setup. Local models, Model service and Cloud API are separate sources. They retain separate form drafts; switching does not automatically generate or upload recordings. Choose the provider for the cloud route; creating a cloud voice requires account permission.

The model-service route is optional for CosyVoice3 and Fish Speech S2 Pro. It requires a running local or remote server with the matching protocol; that server loads the model. The default address is only an example. The experimental route under Local models runs without a separate server and requires a complete audio.cpp GGUF package, not the original checkpoint folder. Actual generation and quality remain unverified.

Click Local GGUF configuration in the service audition form to open local configuration for the matching model. The setup guide provides the same shortcut. It switches the form without copying service credentials or references, preparing a model, generating audio or enabling interpretation.

123

1. 1Choose the family matching your downloaded model.
2. 2Select the complete model directory.
3. 3Choose a clear single-speaker reference.

1234

1. 1Open local GGUF configuration for the matching model without a separate server; generation remains unverified.
2. 2Choose the model loaded by your server.
3. 3Enter the running local or remote server address; the default is only an example.
4. 4Choose a reference; only Generate sends it to the configured service.

Local models and required inputs

12

1. 1Choose VoxCPM2 and check the model and reference recording.
2. 2Load the model ahead of time. Preparation does not generate or play audio.

Prepare and audition Index-TTS2 / 2.5

1. Select the original downloaded model root and click Check model (called Check components and reference in older versions). A preparation notice or missing native files does not mean all original weights are missing.
2. Click Prepare native model. Conversion writes into .boundless-flow/native inside that model directory, retains originals and links existing large weights where possible. The fixed directory is reused rather than timestamped; failures do not trigger an automatic move to C:.
3. Wait for completion and the automatic path switch, then inspect again. New conversion files need space and write permission on the source volume. Do not move original weights while preparation is running.
4. Import a valid reference and enter synthesis text. Generate, audition and download; save the voice or use it for interpretation only after reviewing the result.

123

1. 1Check the original model directory.
2. 2Check the model before preparing it.
3. 3Prepare the native model when conversion is required.

Reference recordings and voice library

Most local cloning routes use 1–60 second references, preferably 3–10 seconds. FireRed requires 1–10.24 seconds, preferably 3–5 seconds of clear single-speaker speech. Use Import reference recording to import a supported audio file. A converted mono 16 kHz WAV and source record are saved while the original is retained. Do not pad a short recording with silence to make it appear valid. The voiceprint workbench has separate input requirements.

Enter the matching reference transcript when required by the model. Search, rename, archive, restore or select a library voice as a reference. Archiving only hides it; it does not delete recordings or free disk space. Applying a reference is an explicit action.

123

1. 1Select the source recording.
2. 2Give the recording a recognizable name.
3. 3Import the recording; the original is retained.

Design, emotion and A/B audition

For voice design, provide voice characteristics and audition text using a model or service supporting design. Emotion parameters, instructions and expression tags apply only when supported by the selected backend. Controlling synthesized emotion is different from detecting emotion in input audio.

Keep local and service results in A/B comparison and use the same text and language where possible. Total request time is not first-audio latency. Download or save session-only comparison results for long-term retention.

123

1. 1Choose voice design and its matching model.
2. 2Describe the voice, pace and delivery.
3. 3Enter the audition text.

VibeVoice Realtime 0.5B service audition

1. Select VibeVoice Realtime 0.5B under Model service. Start the official service separately and enter its actual address, such as ws://127.0.0.1:3000. The app does not download weights or launch Python.
2. Enter an exact preset offered by the service, such as en-Carter_man. The adapter verifies the service voice inventory instead of silently falling back to another preset.
3. Generate and play the audition. Changes to text or settings require regeneration. Only a successful current audition can be applied to continuous interpreting; failed saves can be retried. Cancel or close the voice panel to cancel its audition.

123

1. 1Select VibeVoice under model service.
2. 2Enter the running service address.
3. 3Use an exact preset name from the service.

Inspect model packages awaiting support

MOSS-TTS, FireRedTTS3 Base, dots.tts and other pending models cannot synthesize speech here yet. You can inspect MOSS and FireRed packages for missing files or architecture mismatches. A passing file check does not make a model runnable. See the next section for offline FireRed Instruct audition.

1. In Voice Tools → Model Setup, expand Models being integrated. Choose MOSS-TTS or FireRedTTS3, then select its root or an individual model directory.
2. For MOSS, place the codec in a MOSS-Audio-Tokenizer or MOSS-Audio-Tokenizer-v2 subdirectory of the model root, or select a separate directory in the optional field. If both versions exist, select one explicitly. The checker flags different declared sample rates.
3. Click Inspect package and review the models and dependencies separately. For a missing MOSS Audio Tokenizer, check the codec path. If FireRed weights exist only in downloader staging, finish the download. Do not treat staged files as a final model.
4. Inspect again after changing the path or model. Inspection does not download, convert, enable or save a model configuration, and it leaves the originals unchanged.

1234

1. 1Choose the model family.
2. 2Select the model root or an individual model directory.
3. 3Choose a separate codec directory here when needed.
4. 4Run the read-only inspection.

1

1. 1Check the missing dependencies and path; a file check is not successful synthesis.

123

1. 1Choose FireRedTTS3.
2. 2Select the full model directory.
3. 3Finish the download if only staged weights are found.

Offline FireRedTTS3 Instruct audition

This experimental workflow has not passed full voice-quality or sustained-realtime validation. FireRed now supports experimental generation-time playback in interpretation; see that chapter. Voice design and independent emotion controls remain unavailable. Loading the large model can take substantial time; follow the task status or cancel.

1. In Voice Tools → Model Setup, choose Local models, FireRedTTS3 Instruct and CPU. The current Windows package does not provide FireRed CUDA components. Audition one short sentence to assess speed and voice quality.
2. Select the original model root containing fireredtts3_instruct, redae and text_tokenizer, or select its fireredtts3_instruct subdirectory. This model uses neither the Index weight-preparation workflow nor GGUF.
3. Click Check model. After the check passes, you can select Preload model. This loads the model without generating speech and requires no reference recording or text. The first utterance still needs reference processing.
4. Choose a clear single-speaker WAV lasting 1–10.24 seconds, preferably 3–5, and enter the words actually spoken. After importing or replacing a reference, select Check model again before generating a short audition; do not reuse the old reference check. Compatible mono 24 kHz recordings are retained exactly. Other supported PCM rates produce a derived 24 kHz reference while keeping the original. Choose a shorter recording if the reference exceeds the available context.
5. Enter the synthesis text and generate. Wait for the audio before playing or saving it. Validating audition does not mean generation has completed. Check again after changing the model path or backend.
6. Save the recording or add it to the voice library after listening. The library retains the generated original, derives a mono 16 kHz reference and records generation provenance and version history.

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.

12345

1. 1Choose offline FireRed Instruct audition.
2. 2Choose the original root containing Instruct, RedAE and the tokenizer.
3. 3Optionally preload the model without generating speech.
4. 4Enter the actual reference transcript.
5. 5Enter the synthesis text, then check and generate.

Cloud cloning and voice design

In Voice Tools → Model Setup, choose Cloud API and MiniMax, then enter the endpoint and API key. Check that your account can create voices. The forms below show cloning and design.

1. Clone: import a reference into the voice library, open the cloning tab, select it and enter a voice ID. Follow the service requirements for the recording. The recording is uploaded only when you create the voice.
2. Design: open the design tab, describe the voice, pace and delivery you want, and enter audition text. Voice design uses a description rather than a reference recording.
3. Listen to the result and check the returned voice ID before applying it to speech settings. If a request fails, inspect its status and confirm it did not complete before retrying, to avoid creating duplicates.

12

1. 1Select a reference from the voice library.
2. 2Enter the ID for the cloned voice.

123

1. 1Describe the voice you want to hear.
2. 2Enter the audition text.
3. 3Create the voice after checking the description.

Qwen3-TTS: cloning and voice design

Register models once in Setup Wizard. Choose Base 0.6B or 1.7B for cloning, and 1.7B VoiceDesign for design: these tasks need different weights. In the voice panel, select a configured model instead of entering its folder again. Ordinary cloning and voice design use local native components.

1. Clone: choose a reference from the voice library and check its transcript. It must match what the recording actually says; add it if missing. Enter the new speech, choose Generate, and listen when the audio appears.
2. Design: choose voice design, describe the voice and delivery—for example, an adult woman with a warm voice and moderate pace—and enter the sample text. No reference recording is required.
3. Save a design you like to the voice library with its sample transcript. Later, select a Base model and use that voice as a cloning reference.

The first generation includes model loading and reference processing and is usually slower. Consecutive generations with the same model can reuse loaded components. CustomVoice presets, the automatic design-then-clone workflow and some advanced options still use compatibility components; they are separate from ordinary native cloning.

Qwen cloud audio: which connection generated this recording?

The source labels described here require the Windows r512 or later installer preview.

In Voice Tools → Model Setup → Cloud API → Qwen Audio 3.1 · TTS, choose a connection saved in Setup. The service shown at the top is used for the next generation. After generation, Source of this audio shows the original connection, service host and model used for that recording.

If the saved connection changes, an existing result can show From previous settings. That audio still belongs to the source recorded in the result. You can continue listening, download its WAV, or keep it in A/B; each comparison take retains its own source. A new generation uses the current saved connection.

123

1. 1The connection shown above is used for the next generation.
2. 2The result records its original service and model; From previous settings keeps that origin explicit.
3. 3Continue listening or download this audio. Save needed recordings before changing settings or reloading.

Earlier versions kept playback and A/B copies in the current interface session, so recordings needed downloading before a reload. Updates with the restoration workflow below can recover locally saved audio after this window reloads. Download recordings you need to keep long term.

If the saved service or credentials are unavailable or a connection-change message appears, check and save the connection again in Setup, then reselect it. Wait for cancellation to finish before changing the service. Judge voice and expression by listening; a successful request does not establish cloning accuracy or live-interpretation latency.

Continue listening after finishing Setup

This workflow requires the r480 or later Windows candidate installer. The r475 installer does not include these improvements.

1. After a successful Qwen Audio 3.1 audition, wait for local saving to finish. For comparison, keep a result as A or B; shared A/B accepts local-model, model-service and cloud results, including MP3. Setup will not reload the interface while an audition cache write is pending.
2. Finish and save Setup in the same application window. After the interface reloads, reopen Voice Tools. The latest locally saved Qwen audition and A/B takes are restored when their workbenches open, rather than loading audio at application startup.
3. Check the restored recording's original source before listening. A saved connection change or disabled provider does not reassign the old recording or re-enable that provider. The source label records the original request; it does not verify the current API key. A new generation still needs a currently usable saved connection.
4. Select Play to listen. Recovery does not automatically play, call a provider or upload reference audio. You can continue listening to or downloading the old recording while choosing a different saved source for the next generation.

123

1. 1The currently saved service is for a new generation; it may differ from the recovered recording's source.
2. 2Restored from this device and From previous settings refer to this audio. The source record does not verify that current credentials remain valid.
3. 3Select Play to listen, Download WAV to keep it, or Keep as A / B to compare. Restoration alone performs none of these actions.

Temporary storage and keeping a recording

Download audio you need to keep. This cache supports interface reloads in the current window; it does not provide permanent storage or recovery after shutdown and restart. Browser storage restrictions, manual data clearing, expiration or capacity cleanup can make restoration unavailable. If a local-save warning appears, the current audio remains available for playback and download; download it before reloading or closing.