Find a model for your task, check its supported route, then save its settings once in Setup.
This catalogue describes the routes implemented in Boundless Flow 0.5.0. A listed model is not necessarily bundled, installed or available to your account. Expand a card before downloading or configuring it.
For Chinese transcription, start with SenseVoice. For local cloning, use a prepared Index model, Qwen Base or VoxCPM2. For voice design, choose Qwen VoiceDesign. Test one target language before combining transcription, translation and speech in the interpreting studio.
2Run a short recognition, translation or voice test
3Save it, then select the saved entry in your workbench
Local models: CPU or NVIDIA GPU?
Obtain model files separately. In Setup, check the selected model's components and device before choosing CPU or NVIDIA GPU; device support differs by engine. After the checks pass, generate a short voice sample.
The Windows r469 native CUDA component includes code for compute capabilities 7.5, 8.0, 8.6 and 8.9, plus a PTX forward-compatibility path. VoxCPM2 generation has been verified on RTX 4090. Other GPUs and models still need a short trial; this does not establish universal compatibility. Before loading a model, the app checks the GPU, driver and component. If a check fails, choose CPU in Setup. FireRed currently uses CPU.
For local Windows installation, use the NSIS installer or the r475 locally test-signed MSIX. The old r468/r469 unsigned previews cannot be installed directly. For local MSIX testing, explicitly trust its matching public CER first; this does not establish Microsoft Store certification. Local MSIX installation steps →Installation and first setup →
What the labels mean
Native
A local inference route is implemented. Matching model files and native components are still required.
Compatibility
Uses a dedicated or legacy runtime; some advanced features remain here.
Service
Connects to an already running service or cloud API. Models and permissions come from that service.
Experimental
An entry is implemented, but full product acceptance is incomplete.
File checks only
Checks files or identifies a model; no inference through this entry.
Template only
Saves configuration fields; the corresponding audio protocol is not implemented.
Implemented protocols still need a working endpoint and account permissions. This catalogue does not claim that every provider account has been tested.
No matching entries. Try another model name or clear the filters.
02
Speech to text
4
Native
SenseVoice Small · ONNX
Transcribe Chinese recordings and retain the model's language, emotion and sound-event tags.
Setup & limits
Files
A folder containing model.onnx and either tokens.json or tokens.txt.
Start
Register the ONNX folder in Setup, save it, then select it in recording settings.
Limits
Emotion tags are an aid to review. Unknown stays unknown; tags do not measure intensity.
Transcribe with the native speech component using a single complete model file.
Setup & limits
Files
A SenseVoice.cpp-compatible GGUF file. gguf-fp16-sense-voice-small.bin has been tested; its contents are GGUF despite the .bin suffix.
Runtime
The native-stt component. Select the model file, rather than an original weight folder.
Limits
Requires a component built with the SenseVoice backend. Auto currently defaults toward Chinese; select the source language for fixed-language recordings.
One whisper.cpp GGML ggml-*.bin file compatible with the supplied native component.
Runtime
Select the model file in Setup. This route does not accept arbitrary GGUF files, Hugging Face folders or safetensors.
Limits
Set the recognition language in recording settings. English-only models cannot transcribe Chinese. Available devices depend on the installed component.
Recognize Chinese, English and Japanese recording segments with the local Rust backend.
Setup & limits
Requirements
Requires a build with FunASR enabled and its matching model folder.
Limits
Check the recognition card’s component status. If this build lacks FunASR, use an installer containing it. Keep root model.pt, config.yaml and Qwen3-0.6B configuration and tokenizer; separate Qwen weights are not required. Inspect the files, then try a short recording. This entry is not a native token-streaming recognizer.
Use Base to clone from a reference recording, or VoiceDesign to create a voice from a description.
Setup & limits
Tested variants
Local offline generation has been verified with 12Hz Base 0.6B and 1.7B, and VoiceDesign 1.7B.
Files
Keep the complete model folder and speech_tokenizer subfolder, including their configuration and safetensors weights. Register each variant under a different name.
Start
Register the model in Setup. In Voice Studio, select Base for cloning or VoiceDesign for design; these weights serve different tasks.
Clone a voice from a reference recording, with text-based emotion prompts.
Setup & limits
Prepare
Select the original folder in Setup and prepare the native model. Original files are retained; prepared files live under .boundless-flow/native in that folder. Select the prepared entry when using it.
Model files
The prepared folder contains gpt.safetensors, s2mel.safetensors, acoustic features, semantic codec and matching BigVGAN / W2V-BERT files. Follow the component check for missing files.
Reference and language
Import a reference through the voice library. Native references require mono 16 kHz PCM16 WAV, 1 to 60 seconds. Auto is the current safe language choice for Index.
Some advanced emotion controls remain in the traditional voice panel.
Setup & limits
Limits
Emotion-reference audio and emotion vectors are not all migrated to the shared native interface and may need the compatibility component. Native text emotion prompts use a separate route.
Generate cloned speech from a reference with the native runtime.
Setup & limits
Files
Keep the complete folder with main weights, configuration, tokenizer and audiovae.safetensors. Check or prepare the model in Setup if AudioVAE is missing.
Limits
Register an optional matching Runtime in Setup. This route does not provide a separate emotion-intensity control or VoiceDesign mode.
The local native route needs a matching GGUF package.
Setup & limits
Requirements
Use a complete GGUF package accepted by the current native engine, plus matching components. Original upstream training weights are not a direct substitute.
Limits
The route and file checks are implemented, but full generation-quality acceptance is incomplete. A running model service is an alternative.
Connect to a running local or remote model service and send a reference for a voice test.
Setup & limits
Requirements
The service loads the model. Enter its address, test a voice, then use it for interpreting. The app does not start this service automatically.
Limits
The service cloning adapter is implemented. References are sent to the configured address; the service must match its protocol, not just expose a web page.
Generate speech with preset speakers through the official-style realtime service.
Setup & limits
Protocol
Requires /config and a /stream WebSocket, returning 24 kHz audio.
Limits
The adapter sends complete text and uses preset speakers. Reference cloning, voice design and speed control are not implemented. A file check cannot replace a running service.
Experimental local CPU speech generation. The current Windows installer does not include CUDA runtime components for this engine.
Setup & limits
Requirements
Requires matching Instruct / RedAE models and dedicated native runtime components. Select CPU as the compute device. A downloaded model folder alone is insufficient.
Limits
Separate voice design and emotion controls are not provided, and full product acceptance is incomplete. Start with a short audition using Index, Qwen Base or VoxCPM2.
Register and inspect model files; this route does not yet generate native speech.
Setup & limits
Limits
Folder recognition and file checks for the MOSS family are present. A passed check does not mean inference, cloning or interpreting voice output is supported.
Connect to a compatible translation service, or use the experimental managed local route.
Setup & limits
Model selection
Configure the actual served Index-Translate model ID, such as 2B, 9B or 35B-A3B.
Local route
The experimental managed route needs an accepted GGUF file, correct model metadata and matching service components. Arbitrary original folders will not start.
Limits
Public-service availability and limits depend on its operator. Test translation after saving; saving alone does not verify service availability.
Extract voice embeddings, build profiles and compare recordings.
Setup & limits
Files
Select a single ONNX file compatible with the sherpa-onnx speaker extractor. CAM++, ERes2NetV2 and an embedding.onnx route have been exercised.
Limits
Model format and embedding dimensions must match. A matching filename is insufficient. Similarity assists comparison; it is not identity authentication or a calibrated probability.
Classify emotions in recorded segments for meeting and voiceprint review.
Setup & limits
Files
Use an accepted Base / Seed ONNX package with matching metadata. Revision and structure are checked; an identically named file is not sufficient.
Limits
Provides nine classes. Short, quiet or overlapping speech may be unreliable. Scores do not measure emotion intensity and are not a psychological diagnosis.
Generate speech through a cloud or compatible API.
Setup & limits
Implemented protocol
/audio/speech with PCM output; the default model template is gpt-4o-mini-tts.
Limits
The endpoint must supply the chosen model and voice IDs. A generic compatible endpoint does not automatically support reference cloning or voice design.
Use speech generation plus account-based voice cloning and design.
Setup & limits
Implemented protocol
TTS SSE and bidirectional streaming, with templates including speech-2.8-turbo. Account voice listing, cloning and design routes are implemented.
Limits
Account permissions, models and reference requirements depend on the service. The cloud clone route requires 10 to 300 seconds, rather than the local 1 to 60-second rule.
Use cloud bidirectional speech generation and voice enrollment.
Setup & limits
Implemented protocol
CosyVoice bidirectional TTS and voice enrollment. Enrolled voices must match supported target models.
Limits
Configure cloud CosyVoice separately from a local CosyVoice3 service. Available models, enrollment limits and charges depend on the account and service.
Send an audio file to the configured transcription service.
Setup & limits
Implemented protocol
OpenAI-compatible file transcription and a dedicated qwen-audio-3.1-asr-flash adapter. Qwen file requests are split into chunks of up to 180 seconds.
Limits
File recognition and realtime WebSocket recognition are separate routes. Cloud recognition sends audio; first verify the language and model with a short file.
Choose automatic Chinese/English/dialect recognition or multilingual recognition with separate entitlement.
Setup & limits
Configure once
In Setup → API, save this product’s App ID, API Key and API Secret with the complete /ast/communicate/v1 endpoint. Choose a named mode, save, then select the service in Recording settings. The saved mode is read only there.
Language & previous settings
Chinese/English/dialect mode requires Auto source and empty hints. Multilingual supports 37 documented languages and needs separate permission. Classic RTASR/IAT values remain visible as a previous configuration and require deliberate migration.
Completion & availability
Stop drains the remaining audio and waits for final confirmation; failure does not automatically reconnect or resend. Normal meeting storage retains original audio for retranscription; check the disk if a storage warning appears. Provided in the next source update; the current Windows r475 installer does not include this route. Live-account accuracy and latency are not established.
Stream speech for server segmentation and confirmed text through a connection separate from Ark.
Setup & limits
Configure once
In Setup → API, save openspeech.bytedance.com, a speech-console key and an enabled volc.seedasr.sauc.duration / volc.seedasr.sauc.concurrent resource. Select Realtime API → Doubao when recording to reuse the connection.
Language & completion
Use Automatic source language and empty hints. Server segmentation and final confirmation continue after Stop. Disconnection retains the recording without automatic retry/resend or promoting drafts.
Availability
Included in the r468 candidate. Save the model in Setup, select the service when recording, and use Manage service to change it. Account access, billing, recognition, physical devices and performance remain unverified. Separate from full interpreting or cloning, with no fixed-latency promise.
Stream speech and receive interim text and server-confirmed sentences.
Setup & limits
Setup
Setup → API: save the Beijing endpoint, key and paraformer-realtime-v2. In Recording settings, choose Realtime API → Aliyun · Paraformer and reuse the saved credentials.
How it works
16 kHz mono PCM; approximately 400 ms server pause threshold. Stop still waits for task-finished. On failure, the original recording is kept without automatic reconnection or resend.
Availability & acceptance
Realtime transcription is included in r466; the r468 candidate centralizes model settings so recording selects the saved service. Protocol and local simulated connections were checked; paid-account and physical-microphone acceptance remain pending. Separate from file ASR, cloning and full interpreting.
Continuously send microphone audio and receive translated speech and captions in one target language.
Setup & limits
Implemented interface
The Studio uses the dedicated /v1/realtime/translations interface with gpt-realtime-translate. Source language is automatic, with 13 target languages and optional source captions.
Limits
Requires a saved OpenAI connection and account access. Voice style can follow the source audio; exact voiceprint cloning, fixed voices and custom instructions are not provided. Test speed and quality with your network and language pair.
The qwen-audio-3.1-realtime-plus WebSocket with 16 kHz input and 24 kHz output; Beijing and Singapore region settings are supported.
Limits
Configured separately from the local STT → translation → TTS pipeline. Account permissions, network and available voices affect results; this is not an offline model.
Some entries let you save settings, but their corresponding audio protocols are not implemented.
Setup & limits
Templates only
Classic iFLYTEK RTASR/IAT and TTS, MiniMax ASR, legacy ASR fields under Ark (separate from Doubao Speech), and gpt-realtime-2 entries.
Limits
This entry does not imply generation or recognition support. For audio now, choose an implemented route above; other implemented features from the same provider remain available.
Choose OpenAI, DeepSeek, Kimi, MiniMax, Volcengine or GLM.
Setup & limits
Implemented protocol
Meeting and Linglu text requests use each provider's configured compatible chat endpoint. Enter a valid address, credentials and an actually served model ID.
Limits
These are text-report providers. Listing one does not mean all its audio models are supported. Cloud requests send the text to be summarized.
The app does not download model weights just because you select an entry. Obtain files from the official project, choose a storage drive with enough space, then register their actual location. No automatic model download is triggered by this page.
Provider & mark references
Marks identify project maintainers, service providers or deployment tools. They do not imply partnership or endorsement. Some cards use a clearly labelled text identifier. Open-weight models have individual terms; check the official project before use.
These links identify the projects and providers shown above. They do not extend the app's supported protocols.
In Setup, open Local model library, choose the use and model family, then enter a recognizable name, optional size or quantization, and the existing local folder or file. Recognition, voice, interpreting, speaker and emotion tools select saved entries by name. To get files, use Prepare download for the selected family, review the source and destination disk, then explicitly start. Check files and try a short task before regular use. Finish and save Setup to retain the draft.
App configuration example; no model inference is shown.
Boundless FlowUser manual
