Sprachwerker Studio

Browser-based recording and review studio for TTS and ASR speech datasets. Built for dialect speakers, emotion-aware voice models, and anyone who needs clean, reviewed training data from spontaneous speech.

Supports German and English UI. Runs fully local — no cloud, no external APIs required.

What It Does

Recording

• Records audio directly in the browser
• Dialog mode: a local LLM (Voxtral) generates a prompt, Piper speaks it aloud, you respond — only your microphone is recorded
• Splits speech into sentence-like segments via the space bar
• Runs local ASR (Voxtral) on each segment

Free Recording

• Dedicated tab for uninterrupted free speech — no space bar required
• Record as long as you like, then stop; silence detection automatically splits the audio into segments
• Adjustable minimum silence threshold (default 700 ms)
• Transcribe all segments with one click, review and edit inline, save directly to any dataset profile

Emotion Recording

• Kokoro-compatible emotion labels: neutral, happy, angry, surprised, sad, whispering, question
• Emotion-actor mode: when a target emotion is set, Piper actively triggers that emotion — tells a bad joke for happy, says something provocative for angry, whispers a secret for whispering, and so on
• Designed to capture natural, authentic emotional reactions rather than acted speech

Review

• Manually review, edit, and approve transcripts
• Stores approved items into automatic duration buckets
• Tracks progress toward TTS and ASR collection goals
• Flags duplicate and near-duplicate sentence families

Dataset Management

• Named dataset profiles with per-dataset emotion, dialect, and task labels
• Export-ready manifest format (JSON Lines) with full metadata per segment

Why This Exists

Standard TTS training data tools assume clean studio recordings of read sentences. That fails for:

• Dialect speakers — reading written sentences triggers code-switching to standard language
• Emotion datasets — acted emotions sound unnatural; prompt-triggered reactions sound real
• Spontaneous speech — informal sentence structure, natural pauses, authentic prosody

Sprachwerker solves this by prompting the speaker with a topic or situation, letting them speak freely, then transcribing and reviewing the result.

Storage Model

Approved data is stored under:

output/_sprachwerker/<language>/<dataset-bucket>/

Each saved item produces:

• audio file
• text file
• manifest entry in _review_manifest.jsonl

The manifest format is documented in docs/manifest-schema.md.

Buckets

Saved segments are automatically routed by audio duration into:

• -kurz
• -mittel
• -lang
• -sehr-lang

Start

Prerequisites:

• Docker with Compose support
• ROCm host or CUDA host, depending on backend

Quickstart:

cp .env.example .env

For ROCm, the defaults in .env.example already point to ROCm builder/runtime images.

Start with ROCm:

docker compose -f docker-compose.yml -f docker-compose.rocm.yml up --build -d

Start with CUDA:

docker compose -f docker-compose.yml -f docker-compose.cuda.yml up --build -d

Then open:

• UI: http://127.0.0.1:8095/
• Health: http://127.0.0.1:8095/health

For a step-by-step first start including verification and troubleshooting, see docs/first-start.md.

Automatic Bootstrap

The stack now bootstraps both llama.cpp and Voxtral automatically before asr-api starts.

Bootstrap services:

• llama-builder
• model-fetcher

What they do:

• llama-builder clones and builds llama.cpp
• model-fetcher downloads the required Voxtral GGUF files if they are missing

The results are stored in named Docker volumes:

• llama-bin
• llama-src
• voxtral-models

This means the project no longer depends on a prebuilt host llama.cpp folder or a manually populated host model folder.

Backend Selection

Main variables:

• LLAMA_BACKEND
• ASR_BASE_IMAGE
• LLAMA_BUILDER_BASE_IMAGE
• LLAMA_REPO
• LLAMA_REF
• LLAMA_FORCE_REBUILD
• AMDGPU_TARGETS
• CUDA_DOCKER_ARCH
• LLAMA_CMAKE_EXTRA_ARGS

Typical values:

ROCm:

• LLAMA_BACKEND=rocm
• ASR_BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.4-complete
• LLAMA_BUILDER_BASE_IMAGE=rocm/dev-ubuntu-22.04:6.3.4-complete

CUDA:

• LLAMA_BACKEND=cuda
• ASR_BASE_IMAGE=nvidia/cuda:12.4.1-runtime-ubuntu22.04
• LLAMA_BUILDER_BASE_IMAGE=nvidia/cuda:12.4.1-devel-ubuntu22.04

CPU:

• LLAMA_BACKEND=cpu
• ASR_BASE_IMAGE=ubuntu:22.04
• LLAMA_BUILDER_BASE_IMAGE=ubuntu:22.04

Voxtral Download

Default model source:

• bartowski/mistralai_Voxtral-Small-24B-2507-GGUF
• mistralai_Voxtral-Small-24B-2507-Q4_K_M.gguf
• mmproj-mistralai_Voxtral-Small-24B-2507-f16.gguf

Relevant variables:

• HF_TOKEN optional for public repos, required for gated repos
• VOXTRAL_HF_REPO
• VOXTRAL_MODEL_FILENAME
• VOXTRAL_MMPROJ_FILENAME

Dialog mode variables:

• PIPER_URL
• PIPER_DEFAULT_VOICE
• VOXTRAL_TEXT_CLI_PATH
• VOXTRAL_PROMPT_MAX_TOKENS
• VOXTRAL_PROMPT_TIMEOUT_SECONDS

The default docker compose stack now starts a local Piper service automatically.

It installs de_DE-thorsten-high directly inside the Piper image during Docker build, so Sprachwerker no longer depends on an external Piper stack or a host-mounted voices folder.

Configuration

Important variables:

• CONTAINER_NAME
• PORT
• APP_TITLE
• DIALECT_NAME
• UI_DEFAULT_LANG
• SOURCE_LANGUAGE_CODE
• APP_REGION_GROUP
• SOURCE_TEXT_LABEL
• TARGET_TEXT_LABEL
• DEFAULT_TASK_MODE
• DEFAULT_SPEAKER_PROFILE
• DEFAULT_DIALECT_LABEL
• DEFAULT_EMOTION_LABEL
• TTS_TARGET_HOURS
• ASR_TARGET_HOURS
• TRANSLATION_ENABLED
• OUTPUT_HOST_DIR
• REVIEW_SUBDIR

ASR backend variables are also listed in .env.example.

APP_REGION_GROUP can be set to germany or other. If UI_DEFAULT_LANG is left empty, the UI language falls back to German for germany and English for other.

Manifest Schema

Reviewed dataset items are stored as JSON Lines in _review_manifest.jsonl, with one object per approved segment.

• Human-readable field documentation: docs/manifest-schema.md
• Machine-readable schema: schemas/review_manifest.schema.json

Additional Docs

• Bootstrap and backend setup: docs/bootstrap-and-backends.md
• First start and live bootstrap checks: docs/first-start.md
• Dialog mode and prompt flow: docs/dialog-mode.md
• Free recording mode: docs/free-recording.md

Public Repo Notes

Before publishing or sharing:

• do not commit output/
• do not commit private recordings or reviewed datasets
• do not commit local binary blobs unless their redistribution is clearly allowed
• keep model paths and local infrastructure configurable

The repository includes a .gitignore for the most obvious local artifacts, but you should still review the worktree before any public push.

License

GNU Affero General Public License v3 — see LICENSE.

If you run a modified version as a network service, you must make the source available to users of that service (AGPL §13).