Audio Production

The Audio panel generates speech from your scripts using text-to-speech, handles uploaded voiceovers with timestamp-based splitting, supports sound effect sections, and can layer background music under the voice track.

Prerequisites

• A script in Review or Final status (drafts are blocked)
• A voice configured in Settings or selected per-project

Two Modes

TTS Mode (AI Voice)

Generate speech from script text using one of four TTS providers:

Provider	Voices	Cost	Quality
Edge TTS	300+ Microsoft Neural voices	Free	Good
ElevenLabs	Premium cloned/generated voices	Paid	Excellent
OpenAI	6 voices	Paid	Very Good
OpenedAI Speech	6 voices (local GPU)	Free	Very Good

Any of these carrier voices can optionally be post-processed through RVC Voice Clone to sound like your trained voice model (see Voice Cloning below).

TIP

Edge TTS is the default — no API key needed. OpenedAI Speech is also free if you have a GPU. Enable GPU services via Helm (gpu.enabled: true in values.yaml) or with docker compose -f docker-compose.dev.yml --profile gpu up -d for local dev.

Upload Mode (Recorded Voiceover)

Upload your own recorded voiceover and split it into sections by timestamp:

1. Create a project in Upload mode
2. Upload your voiceover recording (MP3, WAV, OGG, FLAC)
3. Set start/end timestamps (mm:ss) for each section
4. Click Split All — sections are extracted as lossless WAV files
5. Re-upload individual sections if needed (re-record a single section without re-doing everything)

Lossless Splits

Upload splits use WAV format to preserve audio quality. You can replace any section independently without generation loss.

Creating a Project

1. Navigate to Audio > Create
2. Select a script (must be Review or Final status)
3. Choose AI Voice (TTS) or Upload mode
4. If TTS: select a voice and service
5. Click Create Project

The script is automatically parsed into sections.

Section Parsing

The parser uses three strategies (in order):

1. Screenplay format — detects stage directions *(...)* and character dialog **Name:** text
2. Structural markers — splits on ### headings, --- dividers, and **Bold Headers**
3. Paragraph fallback — splits on double newlines if the above produce only 1 section

Stage Directions

Text in *(stage direction)* markers becomes scene context — it's not spoken by TTS but appears as production notes you can show/hide per section.

Sound Effect Sections

Insert SFX/transition sections anywhere in the timeline:

1. Click the + button between sections
2. Choose Sound Effect (instead of Speech)
3. Upload an audio file for the SFX (WAV, MP3, etc.)
4. SFX sections appear with an amber badge and Music icon

Use SFX sections for: transition stingers, sound effects, intro/outro jingles, ambient breaks.

Background Music

Layer background music under the voice track:

1. Expand Background Track below the section timeline
2. Choose Upload or Generate (if MusicGen GPU service is running):
   • Upload — any audio file (MP3, WAV, OGG, FLAC)
   • Generate — describe the music (e.g. "upbeat lo-fi hip hop"), set duration (5-30s), click Generate
3. Configure:
   • Volume — 0-100% (default 15%)
   • Fade In/Out — seconds for gradual volume ramp
   • Mode — Loop (repeats to match voice), Trim (cuts at voice end), Once (plays once)
4. Background is mixed in during Merge

MusicGen

The Generate tab appears when the MusicGen GPU service is running. Enable it via Helm (gpu.enabled: true in values.yaml) or docker compose -f docker-compose.dev.yml --profile gpu up -d for local dev. It creates instrumental tracks from text prompts using Meta's AudioCraft model.

Standalone Music Generation

The Audio > Music panel generates tracks without needing an audio project:

1. Describe the music (e.g. "upbeat lo-fi hip hop, chill vibes")
2. Set duration (5-30s) and click Generate Music
3. Play back generated tracks inline with the play button
4. Download as WAV or Save to Project as background music via the dropdown
5. Use AI Generate to auto-create a prompt based on your channel's topics

Generated tracks persist for the session — you can preview multiple takes and pick the best one.

Generating Audio

• Generate All — processes every pending section sequentially
• Generate One — click refresh on any section to generate just that one
• Regenerate — re-generate with different voice settings

Each section shows its status: pending, generating, ready, or error.

Voice Controls

• Speed — playback speed multiplier (Edge TTS, OpenAI)
• Stability (ElevenLabs) — lower = more expressive, higher = more consistent
• Similarity Boost (ElevenLabs) — how closely to match the original voice

Linking Resources

Click the paperclip icon on any section to link downloaded resources (images, video, audio) from your Resource Library. Linked resources appear as chips — useful for tracking which B-roll goes with which section for later video assembly.

Text Preprocessing

Edge TTS doesn't support SSML, so text is automatically preprocessed:

• Dashes become commas (natural pauses)
• Ellipses become periods (longer pauses)
• Units spelled out: °C becomes "degrees C"
• Abbreviations expanded: e.g. becomes "for example"

Merging

Click Merge to combine all ready sections into a single audio file:

• Configurable silence gaps between sections (0-10000ms)
• Mixed formats supported (WAV + MP3 sections normalized automatically)
• If a background track is configured, it's layered under the voice with fade in/out

Merge with RVC

When your channel has an RVC Voice Model configured in Settings, two additional buttons appear:

• Merge All + RVC — merges all sections, applying RVC voice conversion to each section before concatenation. Sections are copied to temp files, converted through RVC, then concatenated and cleaned up
• Preview RVC — converts just the first section through RVC so you can quickly audition the cloned voice without a full merge

The final output is an MP3 file ready for video production.

Voice Cloning (RVC)

Clone your voice using RVC v2 — it post-processes any TTS carrier voice into your custom voice. RVC is not a standalone TTS service; it's an optional post-processing toggle that works with Edge TTS, ElevenLabs, OpenAI, or OpenedAI Speech. Requires the GPU profile.

Step 1: Collect Voice Samples

Record or gather 10-50 minutes of clean audio of the target voice.

• Avoid background music or noise — consistent mic setup and room acoustics improve results
• Multiple recordings from the same session work best
• Supported formats: WAV, MP3, OGG, FLAC, M4A

Navigate to Audio > Train Voice and drag-and-drop your files into the upload zone. The Setup Guide is located at the top of VoiceTrainPanel, next to the Record Session button. The progress bar tracks total duration — aim for at least 10 minutes (30 minutes ideal).

TIP

You can also record audio externally and upload it via Audio > Create in Upload mode, then export the sections as WAV files for training input.

Step 2: Train a Model

Training happens outside CRAFT using free tools:

• Applio — recommended GUI for RVC training (free, runs locally)
• Google Colab notebooks — free GPU training in the cloud

Use your uploaded voice samples as input. Training takes 20-60 minutes depending on settings (2000-6000 steps). The output is a .pth model file.

Step 3: Upload Your Model

1. In the Upload Voice Model section, select your .pth file (or a .zip containing .pth + .index)
2. The model appears in the Trained Models list
3. Upload multiple models for different voices or styles

Step 4: Enable RVC on Your Channel

1. Open your channel's Settings (gear icon on channel hover)
2. Under Default Voice, select any TTS service (Edge TTS, ElevenLabs, OpenAI, or OpenedAI Speech) and choose a carrier voice
3. Enable the RVC Voice Clone (optional) dropdown and select your trained model
4. Click Save

When you create an audio project, it will use a two-stage pipeline:

1. Your chosen TTS service generates the base speech with the carrier voice
2. RVC converts the carrier audio into your cloned voice

TIP

The carrier voice affects pacing and inflection but not the final timbre — RVC replaces the voice character entirely. Pair it with Edge TTS for a completely free pipeline, or use ElevenLabs/OpenAI for higher-quality carrier speech.

Quick Test

After configuring, go to Audio > Create, select a script, and click Create Project. The service will show your chosen TTS provider plus the RVC model name. Generate a section to hear the result.

Preview

Use Preview Sections to see how your script will be split before creating a project.