Skip to content

Screechback: Local Text-to-Speech Responses

What This Feature Does

Screechback turns a selected text result from a profile-driven dictation into speech and plays it through the Mac. It uses on-device FluidAudio text-to-speech models after their resources have been downloaded.

Typical uses include:

  • hearing an LLM-generated summary without looking at the screen
  • giving a short spoken acknowledgement or response
  • reviewing a processor-stage result while another final result is pasted or saved
  • testing local speech models and voices before assigning them to a profile

Screechback is currently available for interactive hotkey dictation and profiles selected by the local HTTP transcription service. It does not send audio to an Apple Watch, run for Meeting Recording or Transcribe File(s), maintain a conversational thread, or stream speech while text is still being generated.

Where It Runs in a Profile

The profile flow is:

Speech → ASR → Word replacements → LLM graph → selected final output
                                                    ├→ Auto Tag → route matching ─┬→ primary delivery
                                                    │                            └→ ordered routing actions (background)
                                                    ├→ Screechback (background)
                                                    └→ Saved Outputs and Notes (incremental/background)

Screechback is launched as an independent Processing Run branch. Speech synthesis and playback do not delay paste, temporary history, Saved Outputs, Notes, or ordered Trigger Routing actions.

Its current text sources are:

  • Final profile output: the same finalized text that continues to the profile's delivery and storage outputs
  • any configured LLM stage: useful when one branch should be spoken while another stage remains the final profile output

Trigger actions do not currently return text to Screechback. Its available sources are the final profile output and configured LLM stages.

Download and Manage Speech Models

Open App Settings → Local Speech Models & Voices.

This is the app-wide model-management surface. It deliberately calls these speech models, not local language models: they synthesize audio from text and are distinct from the LLMs used for post-processing.

  1. Select a text-to-speech model.
  2. Review what the download contains, its remote FluidInference source, and its local destination.
  3. Click the download button.
  4. Wait for Ready to use.
  5. Use Show in Finder to reveal installed resources.
  6. Open Test Screechback… to compare voices with your own text.

Kokoro downloads its complete published voice catalogue for the selected model rather than requiring one download per voice. PocketTTS and Supertonic-3 also install their included voices or styles with the model.

Selecting a model checks that model's installed state. A download performed in OpenScreech updates the status immediately. If you change model files manually in Finder, select the model again or relaunch OpenScreech to check its current state.

When an installed model is selected, Delete Model & Voices… is available in App Settings and in a profile's Screechback section. After confirmation, OpenScreech unloads the selected model and removes only that model's owned directory, including its installed voices. Sibling model variants and shared parent resources are preserved. The model then reports that a download is required and can be installed again normally.

The first test after downloading or launching can take longer while the selected model loads. This is normal; later tests in the same app session are usually faster.

Download Sources and Runtime Revisions

OpenScreech currently downloads from FluidInference repositories on Hugging Face:

Resources Upstream Repository
Kokoro ANE models and voice catalogues FluidInference/kokoro-82m-coreml
PocketTTS models and voices FluidInference/pocket-tts-coreml
Supertonic-3 model and voice styles FluidInference/supertonic-3-coreml

FluidAudio currently resolves these runtime assets from each repository's main revision. OpenScreech intentionally follows that behavior for now, so a fresh download can receive upstream asset changes without an OpenScreech update. The app shows the actual source before and during download; OpenScreech does not host or redistribute these resources.

The English Kokoro catalogue publishes voices as JSON float arrays. During download, OpenScreech validates each voice's expected dimensions and deterministically writes those same float values into the binary voice-pack layout expected by FluidAudio's ANE runtime. This local packaging step does not train, blend, or invent a voice. The Mandarin catalogue already provides compatible binary packs, so those are downloaded directly.

Choose a Model

Model Catalogue and Output Good Starting Point
Kokoro ANE — English 54 standard voices, 24 kHz WAV Broad English and multilingual-accent voice comparison on Apple Silicon
Kokoro ANE — Mandarin 103 published Mandarin-model voices, 24 kHz WAV Mandarin speech with a large voice catalogue
PocketTTS 6-layer language packs 26 voices, about 450 MB installed, 24 kHz WAV A compact language-specific model with a recommended native voice
PocketTTS 24-layer language packs 26 voices, about 1.3 GB installed, 24 kHz WAV Higher-quality language-specific output when size and generation time matter less
Supertonic-3 — Multilingual 31 languages, 10 female/male voice styles, 44.1 kHz WAV One model for trying several languages

PocketTTS 6-Layer and 24-Layer Packs

The layer count describes the depth of PocketTTS's transformer; it is not a voice count.

  • A 6-layer pack is smaller and faster.
  • A 24-layer pack is larger and slower, but generally offers higher quality.
  • French is currently supplied only as a 24-layer pack.

Start with 6-layer when responsiveness matters, then compare the corresponding 24-layer pack with representative text if quality matters more. Model loading and generation speed depend on the Mac and the length of the response.

Understand Voice Names

The voice picker adds readable catalogue information where it is available.

Kokoro

Kokoro voice prefixes describe language or accent and gender. Examples:

  • af_heart: American English, female
  • am_adam: American English, male
  • bf_emma: British English, female
  • bm_george: British English, male
  • zf_001: Mandarin, female
  • zm_009: Mandarin, male

Other published prefixes identify Spanish, French, Hindi, Italian, Japanese, and Brazilian Portuguese voices. A prefix is useful catalogue metadata, not a guarantee that every voice will sound equally natural for every sentence. Use the tester with the language and style of text you intend to speak.

PocketTTS

Every language pack exposes all 26 included voices. For non-English packs, the picker identifies the voice recorded natively for that language:

Language Native Voice
French estelle
German juergen
Italian giovanni
Portuguese rafael
Spanish lola

The native voice will generally provide the most idiomatic prosody. The other voices remain available for comparison.

Supertonic-3

F1–F5 are female voice styles and M1–M5 are male voice styles. Supertonic also provides a separate language picker. Test the chosen style with the intended language rather than treating the style number as a quality ranking.

Configure a Profile

Open Profiles → Manage Profiles…, select a profile, and find Screechback immediately after Trigger Routing.

Setting Meaning Starting Recommendation
Enable Screechback Generates and plays a spoken response after processing Keep off until the selected model has passed a test
Text source Final profile output or one configured LLM stage Use the summarizer stage for a spoken summary
Speech model One downloaded FluidAudio TTS model Start with the model that matches the response language
Voice One included voice or style Compare realistic text in Test Screechback…
Language Spoken language for Supertonic-3 Match the text supplied by the selected source

Model, source, voice, and language choices remain editable while Screechback is disabled. This lets you prepare or test a profile before enabling playback.

The profile editor also checks and shows the selected model's status, remote download source, local destination, Show in Finder, and confirmed model-deletion action. If the model is missing or has been deleted, it can be downloaded without leaving the profile.

Use an LLM Stage as the Spoken Result

Screechback becomes especially useful when its source is a focused LLM stage:

Original transcript ─┬→ Detailed Formatter → Final profile output → Paste
                     └→ Short Summarizer ───────────────────────→ Screechback

Both LLM branches can begin from the same input. Choose Short Summarizer as Screechback's text source and keep Detailed Formatter as the profile's final output.

Keep spoken-output prompts concise. Text-to-speech will read whatever the processor returns, including headings, symbols, or verbose explanations. A useful summarizer prompt can ask for one or two plain-language sentences with no Markdown.

For a complete preset, see Hear Summaries and Responses.

File-Based Playback and Privacy

Screechback currently synthesizes the complete response as a WAV file and then plays it. It does not begin speaking token by token or sentence by sentence.

The speech synthesis itself runs locally. However, the text may already have passed through a remote LLM post-processor before Screechback receives it. Review the selected processor and endpoint when privacy matters.

Generated playback files are temporary. Normal profile playback removes its WAV after playback or when stopped; the tester keeps its current WAV while Replay is available. Screechback audio is not currently a Saved Outputs or Notes artifact.

On macOS, these WAVs use a dedicated OpenScreech/Screechback folder inside the per-user temporary directory reported under App Settings → Local Speech Models & Voices → Temporary Screechback Audio. That section shows the current file count and size and provides Show in Finder, Refresh, and a confirmed Delete Temporary Audio… action. Manual deletion is restricted to OpenScreech's openscreech-screechback-*.wav filename patterns.

FluidAudio publishes an uncompiled .mlpackage source beside each compiled model component, and renames a component rather than overwriting it when a model is corrected. Neither the sources nor a replaced component is ever loaded, and OpenScreech does not download them again. App Settings → Local Speech Models & Voices → Unused Model Components reports how much they occupy and provides Show in Finder, Refresh, and a confirmed Delete Unused Components… action. A replaced component is only offered once its replacement is present, so nothing is removed that a model still depends on.

Normal profile playback removes its WAV when playback finishes or is stopped. The tester retains only its current WAV while the Replay action is available and removes it when a replacement is generated or the test window closes. Manual cleanup remains available for files left by a crash or forced termination; OpenScreech does not automatically delete temporary audio at launch.

macOS may purge temporary-directory contents while OpenScreech is not running, but it does not guarantee when or whether that cleanup occurs. OpenScreech therefore removes files when their playback lifecycle ends and does not rely on periodic system cleanup.

Troubleshooting

The selected model says it needs a download

  • Download it from the profile or App Settings → Local Speech Models & Voices.
  • If model files were changed manually in Finder, select that model again or relaunch OpenScreech so its installed state is checked again.
  • Confirm the status changes to Ready to use.
  • Review the displayed source and destination before retrying a failed download.

The selected model says an update is available

  • This appears when a FluidAudio upgrade renames a model component. The rest of the install, including every downloaded voice, is kept.
  • The update text names the replaced component. Choosing Update downloads only what is missing rather than the whole model and voice set.
  • After the update completes, OpenScreech removes the superseded component it replaced. Nothing is removed until its replacement is present.
  • FluidAudio 0.15.8 renames the Kokoro tail component, so Kokoro ANE English and Mandarin installs made before that version each report one missing component.

The first test appears slow

  • Allow time for the model to load on its first use after launch or download.
  • Retest a short sentence before diagnosing generation performance.
  • For PocketTTS, compare a 6-layer pack if a 24-layer pack is too slow.

Nothing is spoken after dictation

  • Confirm Screechback is enabled on the profile mapped to the hotkey.
  • Confirm the chosen model reports Ready to use.
  • Confirm the selected LLM stage is still present and enabled when it is the text source.
  • Test the same model, voice, language, and representative text in Test Screechback….
  • Check the Mac's output device and volume.

The voice sounds wrong for the language

  • Confirm the selected text source actually returns the expected language.
  • Choose the native PocketTTS voice for a language-specific pack.
  • For Supertonic-3, set its language separately from its voice style.
  • Compare several voices; catalogue labels describe identity, not a guaranteed quality level.

Current Boundaries

  • Playback is on the Mac only.
  • There is one profile playback channel. A newer Screechback response cancels and replaces speech that is still being generated or played.
  • Each dictation is independent; Screechback does not maintain conversation history.
  • It does not currently speak trigger-action responses.
  • It does not stream partial speech.
  • It does not save generated speech to Saved Outputs or Notes.
  • It runs for interactive hotkey dictation and profiles selected by the local HTTP transcription service, not Meeting Recording or Transcribe File(s).

These boundaries preserve a simple file-based first version while allowing richer response sources, streaming, retained audio, and conversational context to be added later.

See Also