Install and recover models offline

Whisperstream can use model files that you download yourself or copy from another installation. Use this guide if an in-app download fails, a model link stops working, or you want to prepare a new computer without connecting Whisperstream to the internet.

You need one speech-recognition model to dictate. Local enhancement is optional and can use Stream-1 Mini or Customizable. Customizable has two supported Gemma artifacts; you only need one of them.

This guide covers the exact model files supported by Whisperstream. Paths use Windows examples; see Restore models on a Mac for Mac folders and the one model that differs. The PDF and Markdown documentation do not include the model weights. Keep a separate backup of the model files and your installer or source revision and build instructions.

Model download sources#

In-app model downloads use pinned files from Hugging Face. Supported speech-recognition models also have the GitHub backups described below. Existing compatible cached files stay usable. The upstream links in this guide identify the exact supported artifacts for manual recovery.

Automatic download backups#

For Parakeet v2/v3, Qwen and Cohere, Whisperstream first tries the pinned Hugging Face files. If that source remains unavailable after retries, it tries the matching upstream sherpa-onnx GitHub release archive linked below. The app verifies the archive against a checksum shipped with the app before unpacking it, then verifies the required model files before installation. It preserves existing nonempty files. Cancellation, disk errors, checksum mismatches and download-policy failures stop the operation instead of switching sources.

Customizable first downloads Unsloth's pinned Gemma 4 E2B Q4_K_M file. If that source remains unavailable, it automatically tries Google's official pinned QAT Q4_0 file through Hugging Face. These are different artifacts with different checksums and may produce different results. The app verifies the downloaded file and checks the artifact it will actually use. Cancellation, disk errors, checksum or size mismatches and download-policy failures stop the operation instead of switching sources. Stream-1 Mini has no automatic backup. If the automatic sources fail, use the manual recovery steps below. Only the exact supported artifacts listed here are recognized; a similar name does not make another file compatible.

Choose what to restore#

Sizes use decimal MB and GB. Windows may display different numbers. Allow additional space if you download an archive and then extract it. Disk space alone does not determine whether a model will run well on your computer.

For the full language list, see Languages and models.

Find the correct storage folder#

With the default settings, Whisperstream stores models under:

%LOCALAPPDATA%\Whisperstream

Paste that path into File Explorer's address bar. Windows expands %LOCALAPPDATA% to the current user's local application-data folder. Create the folder if it does not exist yet.

Speech models go inside its models folder. Enhancement models go inside local-llm, next to models:

Whisperstream\
  models\
    nemo-parakeet-tdt-0.6b-v2-int8\
    nemo-parakeet-tdt-0.6b-v3-int8\
    qwen3-asr-0.6b-int8\
    cohere-transcribe-03-2026-int8\
  local-llm\
    gemma-4-E2B-it-Q4_K_M.gguf
    gemma-4-E2B_q4_0-it.gguf
    stream-1-mini-vocabulary-q4_k_m.gguf

Only create the folders for models you want to install.

If you changed Model storage location: use the location shown under Home > Advanced > Model storage location as the base instead. For example, if you selected D:\WhisperstreamModels, speech models belong in D:\WhisperstreamModels\models\... and enhancement files belong in D:\WhisperstreamModels\local-llm\.... Do not add another Whisperstream folder beneath that base.

If you use a custom data directory: WHISPERSTREAM_DATA_DIR changes the default base. A saved custom model-storage location still takes precedence. For a normal new installation without either override, use the %LOCALAPPDATA% paths in this guide.

Restore models on a Mac#

On a Mac, the default base folder is ~/Library/Application Support/Whisperstream. In Finder, choose Go > Go to Folder and enter the path. Use the same models and local-llm layout shown above:

~/Library/Application Support/Whisperstream/
  models/
    nemo-parakeet-tdt-0.6b-v2-int8/
    qwen3-asr-0.6b-int8/
    cohere-transcribe-03-2026-int8/
  local-llm/
    gemma-4-E2B-it-Q4_K_M.gguf
    gemma-4-E2B_q4_0-it.gguf
    stream-1-mini-vocabulary-q4_k_m.gguf

Parakeet v2, Qwen3-ASR, Cohere Transcribe, Gemma 4 E2B, and Stream-1 Mini use the same files and checksums as Windows.

Parakeet v3 is different on a Mac. It runs on an Apple silicon engine and uses a Core ML version of the model, not the ONNX bundle in this guide. The ONNX files cannot replace it. Whisperstream downloads it from FluidInference/parakeet-tdt-0.6b-v3-coreml at revision aed02740059203c4a87495924f685de3722ae9ce and checks every file before use. It is stored in:

~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v3/

If you moved it with Model storage location, it is under the folder you chose. To restore it without downloading, copy the complete parakeet-tdt-0.6b-v3 folder from a working Mac's backup into that location. Otherwise, let Whisperstream download it again.

To check a file's SHA256 on a Mac, run shasum -a 256 followed by the file path in Terminal.

Install or restore the files#

1. Quit Whisperstream from its tray menu, or its menu bar icon on a Mac. Closing its window only hides the app. Wait for any model download to stop before copying files.
2. Obtain the matching files. Use your backup, the pinned source linked in the model's section, or the listed upstream archive. You can download on another computer and transfer the files using an external drive.
3. Extract archives before copying. Copy the required files into the exact destination below. Preserve Qwen's tokenizer subfolder. Avoid adding an extra archive-named folder between the destination and the files.
4. Check the files. Compare their sizes and SHA256 checksums with the reference at the end of this guide. Verify every manual copy, including files recovered from an interrupted download.
5. Start Whisperstream. Choose the corresponding language and speech model, or complete Local enhancement setup for the cleanup model you restored.

For a new offline installation, install the app, leave it closed, and place the files before first launch. During onboarding, choose a language supported by the speech model you copied. Automatic selection expects Parakeet v2 for English, Parakeet v3 for its other supported languages, or Qwen for languages outside Parakeet's range. If you intend to use Cohere, finish language setup with the recommended model available, then choose Cohere on Home.

Whisperstream checks the expected filenames in its model folders. It reuses existing nonempty files and downloads missing or empty files when you request a model. It does not search Downloads, scan arbitrary folders, or import an archive you select. Its automatic GitHub backup unpacks only the approved archive through the verified downloader.

A file being detected does not prove it is complete. Existing files are not fully rehashed during ordinary detection. An interrupted file can be nonempty and still fail to load. If that happens, quit the app and replace the affected file with a verified copy. Do not replace a working model merely to fix an unrelated failed download.

NVIDIA Parakeet v2: English dictation#

NVIDIA Parakeet TDT 0.6B v2, converted for sherpa-onnx with INT8 ONNX encoder, decoder and joiner.

Destination folder

%LOCALAPPDATA%\Whisperstream\models\nemo-parakeet-tdt-0.6b-v2-int8\

Required files

encoder.int8.onnx
decoder.int8.onnx
joiner.int8.onnx
tokens.txt

Pinned source: csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8

Revision: 1ab9323565ddb038682214b292f588070a538ce2

Automatic backup and manual recovery source: sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2 on the sherpa-onnx GitHub release. Check the extracted required files against the checksums below before using them.

Use it: On Home, choose English. Automatic selection uses this exact v2 model.

NVIDIA Parakeet v3: multilingual dictation#

NVIDIA Parakeet TDT 0.6B v3, converted for sherpa-onnx with INT8 ONNX encoder, decoder and joiner. It supports 25 languages, including English.

Destination folder

%LOCALAPPDATA%\Whisperstream\models\nemo-parakeet-tdt-0.6b-v3-int8\

Required files

encoder.int8.onnx
decoder.int8.onnx
joiner.int8.onnx
tokens.txt

Pinned source: csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8

Revision: 2bda32ec70b097a55adaa07d9a7173915b43cc78

Automatic backup and manual recovery source: sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2 on the sherpa-onnx GitHub release. Check the extracted required files against the checksums below before using them.

Use it: On Home, choose a supported non-English language. Automatic selection uses v3 for those languages. A new English setup expects v2; copying v3 alone does not satisfy that v2 download check.

Qwen3-ASR: multilingual dictation#

Qwen3-ASR 0.6B, using the sherpa-onnx INT8 export dated 2026-03-25. This is the speech-recognition model, not a Qwen text-generation model.

Destination folder

%LOCALAPPDATA%\Whisperstream\models\qwen3-asr-0.6b-int8\

Required files

conv_frontend.onnx
encoder.int8.onnx
decoder.int8.onnx
tokenizer/vocab.json
tokenizer/merges.txt
tokenizer/tokenizer_config.json

Keep the three tokenizer files inside tokenizer. Do not flatten that subfolder or substitute the older export from a different repository.

Pinned source: csukuangfj2/sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25

Revision: 68818b2313fe77bd06f6a7c5068ff3ef59d02b8a

Automatic backup and manual recovery source: sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25.tar.bz2 on the sherpa-onnx GitHub release. Check the extracted required files against the checksums below before using them.

Use it: On Home, choose a supported language. Automatic selection uses Qwen when Parakeet does not support that language. To use Qwen for another compatible language, open Advanced, turn on Choose speech model manually, and choose Qwen3 ASR.

Cohere Transcribe: optional dictation model#

Cohere Transcribe, using the sherpa-onnx 14-language INT8 ONNX export dated 2026-04-01.

Destination folder

%LOCALAPPDATA%\Whisperstream\models\cohere-transcribe-03-2026-int8\

Required files

encoder.int8.onnx
encoder.int8.onnx.data
decoder.int8.onnx
tokens.txt

Keep encoder.int8.onnx.data beside encoder.int8.onnx. The .data file contains most of the encoder weights and is required. The local folder name contains 03-2026, even though the export source is dated 2026-04-01. Use the destination shown above.

Pinned source: csukuangfj2/sherpa-onnx-cohere-transcribe-14-lang-int8-2026-04-01

Revision: 156a470cf08eefe706a0004f3c52d9ee567ca7a0

Automatic backup and manual recovery source: sherpa-onnx-cohere-transcribe-14-lang-int8-2026-04-01.tar.bz2 on the sherpa-onnx GitHub release. Check the extracted required files against the checksums below before using them.

Use it: On Home, choose a supported Dictation Language, then select Cohere Transcribe in the model picker. This enters Manual selection. Cohere uses the selected language for recognition. Automatic selection does not choose Cohere.

Gemma 4 E2B: Customizable local enhancement#

Google Gemma 4 E2B Instruct (the E2B-it model), using Unsloth's Q4_K_M GGUF. Whisperstream calls it Customizable in the Local model picker.

Destination folder

%LOCALAPPDATA%\Whisperstream\local-llm\

Required files

gemma-4-E2B-it-Q4_K_M.gguf

Download the single Q4_K_M file listed above. Whisperstream uses this model for text only, so no vision projector (mmproj) file is needed. Its runtime and chat template come with the app.

Pinned source: unsloth/gemma-4-E2B-it-GGUF

Revision: 0314792d7f1f7e229411f620751375812bb9faf2

Use it: Open Enhancement, choose Local, and run the speed check for Customizable (Gemma 4 E2B). This file is for text cleanup; it does not replace a speech-recognition model.

Gemma 4 E2B QAT: automatic Google backup#

Google's official Gemma 4 E2B Instruct QAT Q4_0 GGUF is the automatic backup for Customizable when its primary source is unavailable. Its weights, size and checksum differ from Unsloth's Q4_K_M file above, and results may differ. Both appear as the same Customizable choice in the app.

Destination folder

%LOCALAPPDATA%\Whisperstream\local-llm\

Required files

gemma-4-E2B_q4_0-it.gguf

Keep this exact filename. Do not rename it to the Unsloth model's filename or overwrite that model. No vision projector (mmproj) file is needed for text cleanup; the app supplies the runtime and chat template.

Pinned source: google/gemma-4-E2B-it-qat-q4_0-gguf

Revision: 675cff42a74c774d6cb76f76d8eacb49b48c9b93

Use it: Open Enhancement, choose Local, and run the speed check for Customizable. The app checks the Gemma artifact it will use; a result for the other artifact does not qualify this file. An automatic Google fallback remembers that file. If no saved artifact choice exists and both files are present, the app prefers Unsloth. If only one supported file is present, it uses that file. Keep the file you intend to use and verify its checksum; do not overwrite or rename a working model to force a different variant.

Stream-1 Mini: local enhancement#

Whisperstream's own dictation cleanup model, fine-tuned from Qwen3-0.6B-Base. It is the default local enhancement model.

Destination folder

%LOCALAPPDATA%\Whisperstream\local-llm\

Required files

stream-1-mini-vocabulary-q4_k_m.gguf

Keep the exact filename, including lowercase q4_k_m. The vocabulary revision has a distinct filename so an older E model is not reused with new prompts. Do not rename an older Mini file to this name. Plain Qwen3-0.6B is not Stream-1 Mini. Whisperstream supplies the model's built-in cleanup prompts and settings.

Pinned source: whisperstream/stream-1-mini

Revision: ee30dd90bd346d781d0e420d3c6c23244798e7b9

Use it: Open Enhancement and choose Local. The speed check runs for Stream-1 Mini. It supports English Chat and Email cleanup. It does not provide speech recognition or Code-profile enhancement.

If a model link stops working#

Try these options in order:

1. Restore your own backup. A verified copy from a working installation is sufficient. You do not need to register its download with Whisperstream.
2. Look for the exact export or filename. The repository names, revisions and file lists above identify the expected artifacts even if the original page moves. For speech models, the linked sherpa-onnx release archives provide another place to look.
3. Verify a mirror against this guide. Matching SHA256 values establish that the required files are the same bytes as the supported payload. Keep accompanying model licenses and notices with your backup.

A similar model name is not enough. A PyTorch or Safetensors download cannot replace an ONNX bundle or GGUF file. FP16, another quantization, a different model size, a different ONNX conversion, or a newer revision may require app changes. Renaming those files does not make them compatible.

If the exact payload is unavailable, choose another supported model for your language. If only enhancement is affected, leave AI Enhancement off and continue using ordinary dictation. Whisperstream's built-in download links cannot be changed in Settings; manually copying the supported files avoids relying on those links.

Use one local enhancement model#

Stream-1 Mini and Customizable can each be set up alone. Customizable needs one of the supported Gemma files, not both.

1. Copy the GGUF file for the model you want into local-llm.
2. Restart Whisperstream and open Enhancement.
3. Choose Local. Turning on local cleanup checks Stream-1 Mini. To use Customizable instead, choose Customizable in the model picker; Whisperstream checks only that model.
4. Select Start speed check. Whisperstream reuses the file you copied and checks that model on your computer. It does not download the other model.
5. Follow the result to select the model. A passing result lets you use it. A slow-but-usable result requires Use it anyway. A model that is too slow has no override.

Each model needs its own speed check on the computer where you will use it. Restoring a GGUF file does not transfer another computer's performance result. A failed download does not mean your computer is too slow, and one model's result does not qualify the other.

The app supplies the local llama-server engine. If the engine itself is missing or damaged, restore or reinstall the official app. Copying more GGUF files will not repair it. If local cleanup cannot run, turn off AI Enhancement to keep dictating with your installed speech model.

Troubleshoot a restored model#

Verify downloaded files#

Windows PowerShell can calculate SHA256 without installing another tool. For example, to check Stream-1 Mini in the default location:

Get-FileHash -Algorithm SHA256 -LiteralPath "$env:LOCALAPPDATA\Whisperstream\local-llm\stream-1-mini-vocabulary-q4_k_m.gguf"

Compare the returned Hash with the reference below. Uppercase and lowercase letters mean the same thing. Substitute the actual file path when checking another model or a custom storage location. Hashing a large file can take a little time.

The byte sizes below help spot an incomplete download, but matching size alone is not verification. The checksums apply to each extracted file, not to a containing ZIP or tar archive.

Checksums: NVIDIA Parakeet v2#

encoder.int8.onnx (652,184,296 bytes)

a32b12d17bbbc309d0686fbbcc2987b5e9b8333a7da83fa6b089f0a2acd651ab

decoder.int8.onnx (7,257,753 bytes)

b6bb64963457237b900e496ee9994b59294526439fbcc1fecf705b31a15c6b4e

joiner.int8.onnx (1,739,080 bytes)

7946164367946e7f9f29a122407c3252b680dbae9a51343eb2488d057c3c43d2

tokens.txt (9,384 bytes)

ec182b70dd42113aff6c5372c75cac58c952443eb22322f57bbd7f53977d497d

Checksums: NVIDIA Parakeet v3#

encoder.int8.onnx (652,184,281 bytes)

acfc2b4456377e15d04f0243af540b7fe7c992f8d898d751cf134c3a55fd2247

decoder.int8.onnx (11,845,275 bytes)

179e50c43d1a9de79c8a24149a2f9bac6eb5981823f2a2ed88d655b24248db4e

joiner.int8.onnx (6,355,277 bytes)

3164c13fc2821009440d20fcb5fdc78bff28b4db2f8d0f0b329101719c0948b3

tokens.txt (93,939 bytes)

d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d

Checksums: Qwen3-ASR#

conv_frontend.onnx (44,148,281 bytes)

d22dc4423e0940e49884e903d2ea2f7e5567c14fc1aed97e4e26d6b8f208ef9e

encoder.int8.onnx (182,491,662 bytes)

60748d3e6744a57c9c91e1b17424a6c2990567e8adceb0783940c03ed98fa9d9

decoder.int8.onnx (755,914,231 bytes)

4f6885be5959ae26af3089d38ee7972c5fafbeeb1cf8d5e76eab6d8b61ca5771

tokenizer/vocab.json (2,776,833 bytes)

ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910

tokenizer/merges.txt (1,671,853 bytes)

8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5

tokenizer/tokenizer_config.json (12,487 bytes)

4942d005604266809309cabc9f4e9cb89ce855d59b14681fdc0e1cc62ea26c4c

Checksums: Cohere Transcribe#

encoder.int8.onnx (3,090,822 bytes)

cf704f8cfa90e3f0a76f9ffc05998bdf00ba9ae983192c14a85a3a5eb008b367

encoder.int8.onnx.data (2,731,503,072 bytes)

bcf1b7148c8518ae52df1ad2d2fc2b4e89261ea23e6c874eef1d9f55bcbaa4a3

decoder.int8.onnx (153,250,705 bytes)

8372ca6c8ff4db8b916ca3592f5c757a715e691b9edec751ba19b29fc854baf9

tokens.txt (207,437 bytes)

013ede043ae2480e3a9205cc34550d9686100cc682bacc90f702facdfbb93035

Checksums: Gemma 4 E2B#

gemma-4-E2B-it-Q4_K_M.gguf (3,106,738,272 bytes)

740185b21d22ceb83a11c3aa62ad5842ef32c70f6096d756bbee85a1e4ec34b8

Checksums: Gemma 4 E2B QAT (Google)#

gemma-4-E2B_q4_0-it.gguf (3,349,516,256 bytes)

fa401b55b07ee70a54c6dae3903c783a6e65064312529ea57175cb5f8dec6634

Checksums: Stream-1 Mini#

stream-1-mini-vocabulary-q4_k_m.gguf (484,215,648 bytes)

32bef61bdbc98787ac278d96a2aad120a9fc61f588f1f0a4746441e9d6d89ae7

Keep a recovery backup#

After a successful setup, quit Whisperstream and copy the model folders you use to an external drive. Keep them with your installer or source revision, build instructions and this documentation. Back up both models and local-llm if you use local cleanup. On a new computer, restore the files to that computer's model-storage location and complete its own setup and enhancement speed check.