Reading the news with Jutusta v2

Can a synthetic voice read the news well enough that you would want to listen? We had Jutusta v2 read two full ERR news stories: one on the health insurance fund's visit reminders (about 2,700 characters, with quotes and numbers) and a weather forecast (about 1,400 characters, with ranges, units and signs). We picked two voices out of ten and had each read with and without an English delivery instruction, in long paragraphs and short pieces, and the whole article in one request. We made 160 renders in all. Speech recognition checked every version for missing or misread words, and we measured pauses, pace and loudness.

The weather forecast, recommended setup, no instruction

Kerli 1:50
Ahto 1:42

Listen to all 60 clips, with the article text, on the demo page →

What we found

  1. Every version is intelligible. The speech recognition character error rate was 0.3–0.8% on the health fund story and 1.4–2.1% on the weather forecast. Most of the weather errors came from text processing (see point 6). No version was cut short or left text unread.
  2. Long paragraphs and short pieces sound alike. The difference is smaller than between two takes of the same text: one 168-character sentence lasted 9.1–14.5 seconds across takes.
  3. The whole article in one request works just as well as a version put together paragraph by paragraph. The model leaves its own 0.4–1.9 s pauses between paragraphs.
  4. The effect of an instruction depends on the voice. Kerli reads almost the same with or without one. Ahto reads 4–13% faster, with livelier intonation. An instruction adds about a second of synthesis time per paragraph.
  5. Separately synthesized paragraphs differ in loudness. Single-sentence paragraphs came out 5–7 dB quieter. When joining paragraphs, match their loudness.
  6. The errors came from the text, not the voice. Text normalisation read “4-10” as “neli-kümme” and lost the full stop after “14 m/s.”, so two sentences ran together. We fixed it: it now reads “neli kuni kümme” (four to ten), “pluss viis” and “miinus kaks”, and the full stop stays. In the re-synthesized forecast every range and sign was right.

Recommendations for a newsroom

  • Voice. Kerli (a calm, even female voice) or Ahto (a deeper, faster male voice). Ülle, Margit, Kristel, Signe, Veiko, Ragnar, Silver and Tanel also suit the news. Pick by ear.
  • One request per paragraph. A bad take can be redone with another seed without redoing the whole story, and the first audio arrives within seconds: a paragraph typically takes about 6 s, a whole article in one request 17–47 s. Put 0.6–0.8 s of silence between paragraphs and match their loudness. If you would rather not join paragraphs, send the whole article (up to 8,000 characters) in one request; the quality is the same.
  • Instructions are optional and in English. Leave them out by default. Use one when a voice needs a nudge, and keep the wording the same across a whole bulletin so the paragraphs match.
  • Pauses. Write [paus] (about a second) or [pikk paus] where a pause matters, for example before the weather or between stories. Otherwise the model decides the pauses between sentences within a paragraph.
  • Text. Numbers, units, dates, ordinals, ranges and signs are written out as words automatically. Write sports scores (2:1, 0-2), phone numbers and spaced dashes between numbers as they should be heard. Lower-case words typed in capitals for emphasis; acronyms such as NATO are fine. Leave out bylines and photo captions.
  • Reproducibility. The same text, voice and seed give the same audio, so corrections can be made without the voice changing.

Jutusta v2 is available on the web and through the API (model_id: "jutusta_v2"). Every version, the article text and links to the originals are on the demo page.

Sources

The news stories belong to ERR (Estonian Public Broadcasting) and were used for demonstration purposes only. Copyright in the texts remains with ERR; ERR has not produced or endorsed this test.