Make an AI voice read numbers, prices and years correctly: a test list

To make an AI voice read numbers, prices and years correctly, do not guess: put one of each in a short test list, generate it once, and listen. With Sume TTS you can send eight test lines as one job, get a separate audio slice per sentence, and costs under a cent. Where a line reads wrong, respell it in words, such as writing "twenty twenty-six" for a year, and re-test only that line.

The sentence slices, language field and price come from the Sume API reference and API pricing, and the polling steps from Jobs and results, read on 2026-10-03. Sume takes a plain transcript with no SSML and no documented number-format field, so respelling is the control you have. I did not generate audio for this post, so I make no claim about how any voice reads these lines.

What belongs in the test list?

Cover the shapes your scripts use. Keep one idea per line and end each line with a full stop, so sentence segmentation gives one slice per line.

How do I send the list as one job?

Request a wav container, timestamps.words and segmentation with mode: "sentence". Segmentation needs word timings, and per-sentence audio_url slices come only for wav or raw. The script builds the body and prints the cost. After the job finishes, check that segments[] has eight entries; if a decimal or an abbreviation split a line, the count will differ and tell you where to look.

import json
RATE = 0.0475 / 1000
tests = [
    "The price is $1,299.99.",
    "It costs $0.05 per second.",
    "Save 15% before 10/31.",
    "Call 555-0142 today.",
    "Version 2.5 ships in 2026.",
    "The meeting is at 3:30 PM.",
    "Order #4521 left on October 3.",
    "We grew 3x in Q4.",
]
transcript = " ".join(tests)
body = {
    "transcript": transcript,
    "language": "en",
    "voice": {"id": "voi_0123456789abcdef0123456789abcdef"},
    "output_format": {"container": "wav", "sample_rate": 44100, "encoding": "pcm_s16le"},
    "timestamps": {"words": True},
    "segmentation": {"mode": "sentence", "emit_audio": True},
}
print(json.dumps(body)[:120], "...")
print(len(tests), "lines,", len(transcript), "characters =", round(len(transcript) * RATE, 5), "USD")

What does it cost to re-test?

Sume bills $0.0475 per 1,000 characters, spaces and punctuation included, so the 197-character list costs about $0.009. A respelled line is usually longer, so count again before you resend. A test list is cheaper than a retake of a full voiceover with a wrong price in it.

What else should I check?

Listening finds pronunciation errors; it does not prove the words are all there. A round trip through speech-to-text can flag skipped words, as in the STT check. Send language: "en" and an English voice, so the language check cannot stop the request.