> ## Documentation Index
> Fetch the complete documentation index at: https://api-docs.upmore.net/llms.txt
> Use this file to discover all available pages before exploring further.

# Qwen ASR (Ali Bailian)

> Speech-to-text with word-level timestamps through Upmore API — supported models, limits, and why you should extract the audio track before uploading.

Ali Bailian's Qwen ASR models are available through the OpenAI-compatible transcription endpoint. They accept audio **and** video files, and `qwen-audio-3.0-asr-flash` returns word-level timestamps.

## Models

| Model                      | Output                          | Input limit       | Price               |
| -------------------------- | ------------------------------- | ----------------- | ------------------- |
| `qwen-audio-3.0-asr-flash` | Text + sentence/word timestamps | 5 minutes / 2 GB  | ¥0.00022 per second |
| `qwen3-asr-flash`          | Text only (no timestamps)       | 5 minutes / 10 MB | ¥0.00022 per second |

Both models accept audio (`wav`, `mp3`, `m4a`, `aac`, `flac`, `ogg`, `opus`, `amr`, `wma`, …) and video (`mp4`, `mov`, `mkv`, `avi`, `webm`, `flv`, `wmv`). For video input the audio track is transcribed automatically — the model does **not** analyze the picture.

> Timestamps are **word-level, not character-level**: Chinese is segmented into words (`今天`, `给大家介绍`), and each word carries `start`/`end`. The response returns a single `segment` covering the whole file, so split subtitle lines yourself using punctuation and pauses.

## Quick start

```bash cURL theme={null}
curl https://api.upmore.net/v1/audio/transcriptions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F model=qwen-audio-3.0-asr-flash \
  -F response_format=verbose_json \
  -F language=zh \
  -F file=@audio.mp3
```

`response_format` accepts `json`, `text`, and `verbose_json`. Only `verbose_json` carries timestamps.

### Response with word timestamps (`verbose_json`)

```json theme={null}
{
  "task": "transcribe",
  "duration": 46,
  "text": "大家好，今天给大家介绍一款全新的智能早餐机。…",
  "segments": [
    {
      "id": 1,
      "start": 0.12,
      "end": 46.142,
      "text": "大家好，…",
      "words": [
        { "word": "大家好，", "start": 0.12, "end": 0.8 },
        { "word": "今天", "start": 1.16, "end": 1.64 },
        { "word": "给大家介绍", "start": 1.64, "end": 2.8 }
      ]
    }
  ]
}
```

## Extract the audio track before uploading (recommended)

Upmore relays the request to Ali Bailian over the public internet, and the upstream model takes the file as an inline Base64 data URI. Three consequences follow:

1. **Size limit.** Inline Base64 input is capped at 10 MiB — about **7.5 MB of raw file**. A 1080p video is usually far larger.
2. **Bandwidth and cost.** The file travels from your client to Upmore, then from Upmore to Ali as Base64 (\~33% larger than the original). Public egress from the gateway is billed, so a near-limit upload can cost as much in bandwidth as the transcription itself.
3. **Memory.** The gateway buffers the whole file plus its Base64 copy (\~1.33× the file size) per in-flight request.

Passing a public URL would keep the bytes off the gateway, but the upstream then downloads the file itself. Measured on a 538 KB clip: URL input took **21–45 seconds**, versus **3.7 seconds** for an inline upload — slower and unpredictable.

**So strip the audio track first.** A 46-second clip shrinks from 1.4 MB (WAV) to **361 KB** (64 kbps mono MP3):

```bash theme={null}
ffmpeg -i input.mp4 -vn -c:a libmp3lame -b:a 64k -ac 1 audio.mp3
```

| Approach                 | Payload                    | Gateway egress\*                    | Latency (46 s clip) |
| ------------------------ | -------------------------- | ----------------------------------- | ------------------- |
| WAV upload               | 1.4 MB → 1.9 MB Base64     | ≈ ¥0.0015                           | \~5 s               |
| **Extracted MP3 upload** | **361 KB → 481 KB Base64** | **≈ ¥0.0004**                       | **\~5 s**           |
| URL input                | \~0 (upstream downloads)   | ≈ 0 (object-storage egress instead) | 21–45 s, unstable   |

\* At a list price of about ¥0.8/GB. The ASR fee for 46 seconds is ¥0.0101, so an extracted-MP3 upload spends roughly 4% of the ASR fee on bandwidth, while a near-limit WAV or MP4 upload can approach the ASR fee itself.

## Billing

Billing uses the **audio duration** reported by the model: `ceil(seconds) / 60 × 1000` tokens, multiplied by the model ratio. At the list price of ¥0.00022 per second, a 46-second clip costs about ¥0.0101. Video input is billed by its audio duration only.

## Limits and notes

* **5 minutes per request.** Longer files must be split into ≤5-minute chunks; concatenate the results and offset each chunk's timestamps by its start time. (The asynchronous `filetrans` API that accepts 12-hour files is not exposed yet.)
* **File upload only.** Passing a URL as the `file` value is not supported today.
* **Speech only.** Video frames are not analyzed. For a visual timeline, use a video-understanding model such as `doubao-seed-2-1-pro` and merge its timestamps with the transcript.