ElevenLabs Chinese Narration, Tested: Voice and Verification on a 591-Character Long Script

The question everyone asks about Chinese AI voiceover is: “Does it misread characters on long scripts?” Rather than trade opinions, we verified a shipped video: the narration of Veiled Lore Atlas’s Minxiong Haunted House EP02 — 591 characters, recorded in one pass with ElevenLabs’ Ethan Zhang voice, then transcribed and compared against the locked script character by character. The result, and the verification method, are both in this article.

ENVIRONMENT

Veiled Lore Atlas, Minxiong Haunted House EP02 V2 (published to the Facebook Page as a Reel on 2026-10-14)

TOOL

ElevenLabs, voice “Ethan Zhang”; script in Simplified Chinese, 591 characters including punctuation, 10 segments

OUTPUT

a 160.84-second narration file, about 220 characters per minute (591 characters ÷ 160.84 seconds)

VERIFICATION

faster-whisper (large-v3-turbo) transcription plus a normalized character-by-character diff; all 10 segments ASR_VERIFIED

What exactly was tested?

The script was not casual prose. It was written from archive file LOC-0011 (the fact-checked Minxiong Haunted House dossier): the facts were graded and settled first, then written into a 591-character narration in a cinematic suspense-documentary register — no storyteller persona, Simplified Chinese throughout (the series’ standing spec), locked before anyone pressed record. The recording was made in ElevenLabs with Ethan Zhang, a Mandarin male voice, and came out as one continuous 160.84-second narration file.

The order matters: lock the script, then record, then verify character by character. If the script is still moving, verification has no reference. Once the script was approved and locked, the recording had exactly one job — say precisely those characters — and verification checks exactly that.

How does character-by-character verification work — and what did it find?

The method is machine diff plus human adjudication, and neither half works alone:

1. Transcribe the finished audio with faster-whisper (large-v3-turbo).
2. Diff the transcript against the locked script after normalization (punctuation and numeral styles unified first), with an alignment confidence per segment.
3. Adjudicate every difference by listening: was it the recording misreading, or the recognizer mishearing?

The result: all 10 segments correspond — no missing sentences, no additions, no reordering — with per-segment alignment confidence from 0.875 to 1.000. A sample of the segment table (the full timestamp table lives in the production file sentence_timestamps.csv, and the edit aligns shots directly against it):

(The remaining segments — N03, N05, N06, N09 — also passed, at 0.96–1.00 confidence. The narrative spans 0.00–160.02 s, with about 0.8 s of tail.)

Why were half the flagged “errors” false alarms?

The diff did flag a list of differences — and on listening, every one of them was the recognizer writing the wrong homophone while the recording said the right word:

This table is the part of the article most worth stealing: when you verify TTS with ASR, the diff is not a grade sheet. Mandarin is dense with homophones, and recognizers will trip on them by design. Skip adjudication and you will either reject a perfect take, or — worse — rubber-stamp the process until a real misread sails through. Our rule: every difference gets adjudicated against the original audio, the ruling goes into the QA record, and only then does the verdict exist.

The reverse case showed up too: during verification, a phone-generated transcript rendered the script’s “2022” as “2020”, briefly making it look like the recording had misread the year. On re-listening, the recording says 2022; the transcript was wrong. Verification tools make mistakes — treat that as a constant of the workflow, not an exception.

Where are the real risk points in long Chinese scripts?

591 for 591 does not mean Chinese TTS has no mines. This script happened to step on most of them, so here is the map for anyone recording long-form:

• Years and numbers. The script contains four years (1949, 1929, 1945, 2022) and one measured quantity, “1200 ping” (坪, the Taiwanese floor-area unit). Numbers are where TTS most easily picks the wrong reading form — digit-by-digit versus whole-number, and which measure word follows. Spot-check these first.
• Proper nouns and rare words. Liu Family Old House （劉家古厝）, pavilions （亭榭）, Minxiong, Chiayi — uncommon terms and place names are the second minefield; 古厝 and 亭榭 are exactly the words the recognizer fumbled and a human had to confirm by ear.
• Breath on long segments. The longest segment runs 22.14 seconds (N04) at a high information density. Ethan held a documentary pace of about 220 characters per minute across the whole piece, with no speeding up or mumbling at segment tails — that last clause is a human listening judgment, and it’s labeled as one.

Failures and limits: what this test does not cover

House rule — this section is never skipped:

• One voice, one genre. Ethan Zhang × suspense-documentary narration. Other voices, female voices, storyteller or advertising registers are outside this result.
• The script was written to be readable. The lock-first workflow also means the script was smoothed before recording — no pile-ups of polyphonic rare characters, no English abbreviations or model numbers. A script full of technical terms is a different difficulty class, untested here.
• Cost was not logged. This recording left no record of the ElevenLabs plan or character spend, and house rule is not to invent figures. A per-episode narration cost needs a separate check against the then-current plan.
• Verification can’t hear “wrong emotion”. A character-by-character diff proves the words were said correctly, not that they were said well. Whether the delivery fits the picture still takes a human listening to the whole piece — that step has no automation.

Sources and further reading

• Minxiong Haunted House EP02 V2 Source QA and conflict list (2026-10-14): the verbatim verification method, the 10-segment verdict and the ASR difference list are from here.
• Sentence timestamp table sentence_timestamps.csv (Veiled Lore Atlas production files): per-segment start/end times, durations and alignment confidence are from here.
• Locked narration script narration_final_zh-CN.txt and its source dossier LOC-0011: the full 591-character script and its evidence grading are from here.

FAQ