Korean broadcast audio: denoising cut hallucinations but lowered language-detection accuracy (95.2% → 93.4%)

[sungbin1015]

We run large-v3 on Korean TV broadcast audio (news, documentary, drama, period drama, variety, sports — one episode each, about 6h20m in total). Sharing two things that were not obvious to us, in case they save someone time.

1. Denoising helps ASR but hurts language detection

We denoise with DeepFilterNet v3 (atten_lim_db=-30) because broadcast audio always has BGM, effects and crowd noise, and it does reduce hallucinated lines such as ご視聴ありがとうございました in non-speech regions.

We assumed the same denoised audio would also be better for detect_language. It was not. Same VAD segments (Silero, computed once on raw audio and reused for both inputs), faster-whisper-large-v3-turbo, 3,115 speech segments:

Content	Segments	raw	denoised
Baseball	1092	1073	1072
Documentary	383	349	349
Drama	314	282	279
Variety	502	464	438
Period drama	437	411	386
News	387	385	384
Total	3115	2964 (95.2%)	2908 (93.4%)

The drop is concentrated in variety and period drama, the two genres with the heaviest music. There is no public ground truth for this audio; the per-segment review sheet is linked from the report below, so please read the numbers as indicative.

So we split the inputs: VAD and language detection read raw audio, only ASR reads denoised audio.

2. The gates we ended up with

In order of how much they helped:

1. VAD before ASR, so silence and BGM never reach the model. By far the biggest win.
2. Drop segments with avg_logprob < -1.0.
3. If the detected language is not Korean and its probability is below 0.5, force Korean.
4. For utterances shorter than 3 s detected as non-Korean, transcribe twice (Korean and the detected language) and keep the result with the higher mean avg_logprob. Drop it if both are below -0.6.
5. If the language is Korean but less than 30% of the alphanumeric characters are Hangul, drop the segment. This catches kana/hanzi output in Korean mode.

Repetition settings: condition_on_previous_text=False, repetition_penalty=1.2, no_repeat_ngram_size=3.

With BatchedInferencePipeline (batch_size=16) on an RTX 4090 the RTF was about 0.042.

Code and details

• LID comparison tool: https://github.com/SceneMakerAI/poc-lid-bench
• STT benchmark (transcripts and scores in output/): https://github.com/SceneMakerAI/poc-stt-bench
• Report (Korean): https://doc.scenemaker.solbox.com/docs/poc/audio-bench/2

Has anyone measured the same effect with other denoisers? I would also like to know if there is a better approach than the dual-transcribe step for very short utterances.