Power BI Walkthrough — cloned German voice
Repeatable flow for finance reports: check the script, clone the voice, generate narration, mux an MP4.
Model: Qwen/Qwen3-TTS-12Hz-1.7B-Base · Spoken language: German · Software terms stay English.
Client test scenario
Runs on the public NordFinance demo only (no client data). It checks terminology, optionally clones a public demo speaker, reads a short German Power BI script, and muxes an MP4.
Tick Include voice clone on the hosted Space (needs ZeroGPU, about 1–3 minutes). Untick it to test terminology + FFmpeg only.
The checker replaces forbidden translations (Datenschnitt → Slicer, QuickInfo → Tooltip) and keeps finance terms in German.
What you need: 15–30 seconds of clean German, WAV or MP3, 44.1 or 48 kHz, no music. Minimum 3 seconds. Include the exact transcript. Reuse the same voice file for every later video.
Listen to a preview (two chunks) first. If the voice is right, generate the full narration. A 4–6 minute script often takes 3–8 minutes of ZeroGPU time.
Upload a screen recording with no client data. The original soundtrack is replaced. If the recording is shorter than the voice, FFmpeg freezes the last frame.
Procedure for an assistant (no video experience needed)
- Tab Test: click Run client test with voice clone ticked. You should see terminology PASS, German speech with Slicer / Tooltip, and a short MP4. This uses a public demo speaker, not the consultant.
- Get the approved script from the consultant as a
.txt. Swap the terminology CSV only when a new list is sent. - Tab 1 Script: paste the script → Check script. Do not ignore blocked terms. Continue only when the check says it passed.
- Tab 2 Voice: once, upload the reference WAV plus transcript → Create voice → download the
.ptintovoices/. From video two onward, only load that.pt. - Tab 3 Narration: tick Preview only → generate → listen for about 20 seconds. If it sounds right, untick preview → generate the full narration.
- Tab 4 Video: upload a 1080p screen recording with no client data, untick the demo-video box, Create MP4. Save the file.
- Upload the MP4 as unlisted YouTube or Streamable. Put the link in the handover email.
Reference recording — what the consultant must supply
- Duration: 15–30 seconds (minimum 3, 20 is better)
- Format: WAV, 16-bit or 24-bit, 44.1 or 48 kHz, mono or stereo. MP3 works; WAV is better.
- Content: calm High German, one connected block of speech, no laughter, no music, no echo.
- Transcript: word for word, including audible fillers.
Stack and monthly cost
Hugging Face Space (ZeroGPU) + Qwen3-TTS-12Hz-1.7B-Base + Gradio + FFmpeg. The Base model clones the voice from a few seconds and speaks German, so you do not record narration for each video. Gradio runs in the browser. FFmpeg combines voice and screen recording into an MP4. Monthly cost: 9 USD Hugging Face PRO (40 minutes GPU/day, highest queue). ZeroGPU hardware itself is 0 USD. YouTube unlisted / Streamable for the client link: 0 USD. About 9 USD/month, under 40 USD. An always-on T4 Space (0.40 USD/h) would be ~288 USD/month and is unnecessary because one video only uses a few GPU minutes.
Time
First video, including voice setup and glossary check: 2.5–4 hours. From the tenth video, same procedure: 25–40 minutes (trim recording, check script, generate, listen once, mux, send the link).
Proof of cloned non-English TTS
Public demo of the same model: Qwen3-TTS Space — Voice Clone (Base) tab, language e.g. German. Tool: Qwen3-TTS-12Hz-1.7B-Base.
No client data
Use a demo report, masked exports, or the built-in sample page. No VPN, no client tenant.