Hacktheon Sejong 2026 Quals Voice Over Writeup
Hacktheon Sejong 2026 Quals Voice Over WriteupHacktheon Sejong 2026 Quals Voice Over WriteupHacktheon Sejong 2026 Quals Voice Over 문제 풀이Hacktheon Sejong 2026 Quals Voice Over writeup
Contents
Summary
서버는 업로드한 wav에서 text similarity와 speaker similarity를 함께 본다. target 문장은 TTS로 맞추고 reference 음성은 reverse한 뒤 낮은 볼륨으로 뒤에 붙여 speaker embedding만 보태는 방식으로 두 threshold를 넘겼다.
Analysis
검증 기준은 두 값이다.
text_similarity >= 0.8
speaker_similarity >= 0.8일반 TTS로 target sentence를 합성하면 text similarity는 충분하지만 speaker similarity가 낮다. 반대로 reference wav를 그대로 넣으면 speaker similarity는 높아도 transcript가 달라 text similarity가 떨어진다.
그래서 reference 음성이 ASR에는 의미 없는 소리로 들리고 speaker embedding에는 화자 특징이 남도록 처리했다. reference wav를 reverse한 뒤 낮은 볼륨으로 target TTS 뒤에 붙이면 앞쪽 TTS 때문에 ASR은 목표 문장을 주로 인식한다. 뒤쪽 reverse reference는 transcript에는 거의 섞이지 않지만 speaker verification에는 영향을 준다.
sample_003 reverse 기준으로 볼륨은 0.12 정도가 적당했다. 너무 낮으면 speaker similarity가 부족하고 너무 높으면 ASR transcript가 오염된다. 성공한 제출에서는 speaker similarity 0.8026, text similarity 0.8859가 나왔다.
Solver
curl -sS http://3.37.31.209:8000/api/challenge > challenge.json
target=$(jq -r .target_sentence challenge.json)
token=$(jq -r .token challenge.json)
espeak-ng -v en-us -s 135 -w tts.wav "$target"
ffmpeg -y -loglevel error -i tts.wav -ar 16000 -ac 1 tts_16k.wav
ffmpeg -y -loglevel error -i sample_003.wav -af areverse sample_003_rev.wav
ffmpeg -y -loglevel error -i sample_003_rev.wav -filter:a "volume=0.12" ref.wav
printf "file '%s'\nfile '%s'\n" "$PWD/tts_16k.wav" "$PWD/ref.wav" > concat.txt
ffmpeg -y -loglevel error -f concat -safe 0 -i concat.txt -c copy submit.wav
curl -sS -F audio=@submit.wav -F token="$token" \
http://3.37.31.209:8000/api/verify | jq .Flag
hacktheon2026{b7d30e21e4106a6ca4d451a218f15a97}
Summary
The server scores an uploaded wav on both text similarity and speaker similarity. I matched the target sentence with TTS, then reversed the reference voice and appended it at low volume so it only contributed to the speaker embedding. That got both scores over the threshold.
Analysis
The check uses two values.
text_similarity >= 0.8
speaker_similarity >= 0.8Synthesizing the target sentence with plain TTS gives enough text similarity but low speaker similarity. Submitting the reference wav as-is gives high speaker similarity, but the transcript differs, so text similarity drops.
So I made the reference voice sound meaningless to ASR while keeping its speaker characteristics for the speaker embedding. With the reversed reference wav appended at low volume after the target TTS, ASR mostly recognizes the target sentence from the TTS at the front. The reversed reference at the end barely leaks into the transcript but still affects speaker verification.
With sample_003 reversed, a volume of about 0.12 worked well. Too low and speaker similarity falls short; too high and the ASR transcript gets polluted. The successful submission scored speaker similarity 0.8026 and text similarity 0.8859.
Solver
curl -sS http://3.37.31.209:8000/api/challenge > challenge.json
target=$(jq -r .target_sentence challenge.json)
token=$(jq -r .token challenge.json)
espeak-ng -v en-us -s 135 -w tts.wav "$target"
ffmpeg -y -loglevel error -i tts.wav -ar 16000 -ac 1 tts_16k.wav
ffmpeg -y -loglevel error -i sample_003.wav -af areverse sample_003_rev.wav
ffmpeg -y -loglevel error -i sample_003_rev.wav -filter:a "volume=0.12" ref.wav
printf "file '%s'\nfile '%s'\n" "$PWD/tts_16k.wav" "$PWD/ref.wav" > concat.txt
ffmpeg -y -loglevel error -f concat -safe 0 -i concat.txt -c copy submit.wav
curl -sS -F audio=@submit.wav -F token="$token" \
http://3.37.31.209:8000/api/verify | jq .Flag
hacktheon2026{b7d30e21e4106a6ca4d451a218f15a97}