Bug: `response_format=verbose_json` returns empty text for Whisper transcription

[KwinyarutP]

Note: I originally tried to report this issue via email, but the message bounced with "550 5.1.1 Address does not exist." I also could not find a working contact listed on the Docker Hub page (https://hub.docker.com/r/hanseware/vllm-audio-video). Therefore, I am opening this issue here in hanseware/FastFusion instead. Sorry if this repository is not the correct place to report it, and please let me know if it should be redirected elsewhere.

Environment

• Image: hanseware/vllm-audio-video:latest
• Tested on: 2026-03-07, 2026-03-12
• vLLM versions: 0.16.0, 0.17.1
• Model: typhoon-ai/typhoon-isan-asr-whisper
• Endpoint: POST /v1/audio/transcriptions

Description

When using response_format=verbose_json, the API returns empty transcription results.

However, response_format=json works correctly and returns the expected text.

Reproduction

import requests, io, wave, struct, math

RATE = 16000
samples = [int(32767 * math.sin(2 * math.pi * 440 * t / RATE)) for t in range(RATE * 2)]
buf = io.BytesIO()

with wave.open(buf, "wb") as wf:
    wf.setnchannels(1)
    wf.setsampwidth(2)
    wf.setframerate(RATE)
    wf.writeframes(struct.pack(f"<{len(samples)}h", *samples))

r = requests.post(
    "http://<your-server>/v1/audio/transcriptions",
    files={"file": ("test.wav", buf, "audio/wav")},
    data={
        "model": "typhoon-ai/typhoon-isan-asr-whisper",
        "response_format": "verbose_json",
        "language": "th"
    },
)

print(r.json())

Expected result

{
  "text": "...",
  "segments": [...],
  "words": [...],
  "duration": 2.0
}

Segments should include metadata such as:

• avg_logprob
• no_speech_prob
• compression_ratio

Actual result

{
  "text": "",
  "segments": [],
  "words": null,
  "duration": 2.0
}

Notes

• response_format=json returns correct transcription.
• Only verbose_json appears broken.
• This worked in an older build of the Docker image.
• Confidence metadata is required for transcription quality filtering.