The LAS speech-to-text (Doubao-Seed-2.0-lite) enhanced operator provides multimodal refinement capabilities on top of the basic ASR recognition results: it first completes input download, audio extraction, duration detection, and basic ASR recognition. When the routing, language, and basic ASR results meet the refinement conditions, audio slices, basic ASR recognition drafts, and reference hotwords extracted from the hotword context are sent to the multimodal model to correct proper nouns, numbers, punctuation, and some recognition errors. If the model determines that there are missed recognitions, it attempts to fill in missing content.

Operator ID: las_asr_seed-2-0-lite

Application scenarios #

• Meetings and interviews transcription: Asynchronously recognizes meetings and interview audio as text and refines it, improving the accuracy of proper nouns and terminology, automatically segments, and enhances recording efficiency.
• Podcast and course transcription: Provides high-quality transcription and refinement for long audio, suitable for content archiving and repurposing.
• Audio and video subtitles: Automatically converts speech in audio and video into time-stamped subtitles and improves readability through refinement.
• Content review and quality inspection: Recognizes recordings as text, refines them, and then analyzes them based on quality inspection rules.

Core capabilities #

• Multimodal refinement:
  • On top of the basic recognition results, the Doubao-Seed-2.0-lite multimodal model is invoked to correct and supplement proper nouns, numbers, punctuation, and semantic coherence, significantly improving the quality of the transcribed text.
  • If refinement fails, it can fall back to the basic ASR results to ensure usability.
• Multi-format audio and video input recognition:
  • Supports audio and video file input, automatically extracts the video audio track for recognition.
• Comprehensive recognition capabilities:
  • Supports text normalization (ITN), automatic punctuation, semantic smoothing (DDC), and speaker separation.
• Hotwords and sensitive words:
  • Supports direct hotword injection and hotword list enhancement, with hotwords simultaneously sent to the underlying ASR and refinement prompt.
  • Supports sensitive word protection: System-masked content can be excluded from the refinement model, and custom blanking or asterisk replacement is supported.

Supported regions #

Supported languages #

Operator performance #

Input requirements #

Output requirements #

• Billing standards

  Breakdown	Billing standards description
  Billing items	When using the current operator, the refinement model is used by default to refine the speech recognition results. Therefore, the billing items include speech recognition and result refinement: Speech recognition: Billed based on the duration of the input audio/video. Unit: CNY/hour Result refinement: Billed based on the token usage of the refinement model (Doubao-Seed-2.0-lite). Unit: CNY/million tokens. The final cost is the sum of the two fees.
  Billing type	Usage-based billing, with charges billed hourly based on actual usage.
  Unit price	Depends on the model used.

• Billing details
  Billing formula: Total cost = speech recognition fee + refinement fee = (speech recognition unit price * audio duration) + (refinement tier unit price * model usage)

  • Speech recognition unit price:

    Speech recognition model	Unit price
    Seed-ASR 2.0 Enhanced	1.6 CNY/hour

  • Refinement unit price:

    Refinement model	Model input length (thousand tokens)	Cache storage	Text input	Audio input	Text cache hit	Audio cache hit	Output	Unit
    Seed-2.0 Lite	[0, 32]	0.017	0.6	9	0.12	1.8	3.6	CNY/million tokens
    	(32, 128]	0.017	0.9	13.5	0.18	2.7	5.4	CNY/million tokens
    	(128, 256]	0.017	1.8	27	0.36	5.4	10.8	CNY/million tokens

• Billing example

  • Example scenario: A 3-minute video is input, and the current operator is used to convert speech to text. The processing result displays the following model usage.

    "token_usages": [
          {
            "task_info": "refine_chunk_001",
            "model_name": "doubao-seed-2-0-lite-260428",
            "token_usage": {
              "prompt_tokens": 2015,
              "completion_tokens": 6429,
              "total_tokens": 8444,
              "cached_tokens": 0,
              "audio_tokens": 814,
              "audio_cached_tokens": 0,
              "reasoning_tokens": 5599
            }
          }
    

  • Cost details:

    • Speech recognition fee:
      Speech recognition fee = 1.6 CNY/hour * (3/60) hours = 0.08 CNY
    • Refinement fee:
      • Model input token usage (prompt_tokens): 2015, within the [0, 32] thousand token tier.
      • All cache-related tokens are 0, so no cache-related fees are involved.
      • Text input token usage (prompt_tokens - audio_tokens): 2015 - 814 = 1198.
      • Audio input token usage (audio_tokens): 814.
      • Output token usage (completion_tokens): 6429.
        Therefore:

      Refinement cost = text input cost + audio input cost + output cost 
              = 0.6 CNY/million tokens * （1198/1000,000) +9 CNY/million tokens * （814/1000,000) + 3.6 CNY/million tokens * （6429/1000,000)
              ≈ 0.0007 + 6.6332 + 0.231
              ≈ 6.8649 CNY
      

    Total cost = speech recognition fee + post-editing fee = 0.08 CNY + 6.8649 CNY = 6.9449 CNY

LAS provides you with the capability for an "online experience" and offers a certain amount of free trial quota. No configuration is required to experience the data processing results of the LAS operator online.

The enhanced version of the speech-to-text (Doubao-Seed-2.0-lite) operator is an asynchronous interface, with the process divided into two stages: submitting a task and querying results. You need to first create a speech recognition task, and then query the recognition result using the task ID.

API usage instructions #

• The API of this operator is an asynchronous task interface. You need to submit a task through the Submit Task interface to obtain the task's task_id, and then use the Query Result interface to obtain the running status of the corresponding task (task_id) and retrieve the operator's processing result.
• Caution: After submitting a task, the task_id is valid for 3 days. After 3 days, you will not be able to use this task_id to obtain the operator's processing result.

Submit a task #

Interface description #

Submit the audio link and obtain the task ID assigned by the server. This operator enables Doubao-Seed-2.0-lite multimodal refinement by default, and always uses the seedasr resource at the underlying layer.

Request parameters #

Response parameters #

Example #

Request example

curl --location "https://operator.las.ap-southeast-1.volces.com/api/v1/submit" \
--header "Content-Type: application/json" \
--header "Authorization: Bearer $LAS_API_KEY" \
--data '
{
    "operator_id": "las_asr_seed-2-0-lite",
    "operator_version": "v1",
    "data": {
        "audio": {
            "url": "https://${LAS_TOS_BUCKET}.tos-${LAS_REGION}.volces.com/operator_cards_serving/public/${LAS_ENVIRONMENT}/las_asr/badaling.wav",
            "language": "zh",
            "format": "wav"
        },
        "request": {
            "enable_punc": true,
            "enable_speaker_info": true,
            "corpus": {
                "context": "{\"hotwords\":[{\"word\":\"火山引擎\"}]}"
            }
        }
    }
}'

Response example

{
  "metadata": {
    "task_id": "xxxxx123ef24ea40546c",
    "task_status": "PENDING",
    "business_code": "0",
    "error_msg": "",
    "request_id": "494022a8a0fc3eadb758cf8b0e8b20ef"
  }
}

Error code #

Query results #

Interface description #

Query the speech transcription and refinement results using the task ID.

Request parameters #

Response parameters #

Example #

Request example

curl --location "https://operator.las.ap-southeast-1.volces.com/api/v1/poll" \
--header "Content-Type: application/json" \
--header "Authorization: Bearer $LAS_API_KEY" \
--data '
{
    "operator_id": "las_asr_seed-2-0-lite",
    "operator_version": "v1",
    "task_id": "xxxxx123ef24ea40546c"
}'

Response example

{
    "metadata": {
        "task_id": "xxxxx123ef24ea40546c",
        "task_status": "COMPLETED",
        "submit_time": "2026-07-21T16:04:06+08:00",
        "end_time": "2026-07-21T16:04:06+08:00",
        "business_code": "0",
        "error_msg": "",
        "request_id": "d204c21f5c7c8f8cfeb85d211b9c20ac"
    },
    "data": {
        "audio_info": {
            "duration": 3575,
            "language": "zh"
        },
        "result": {
            "text": "参观八达岭长城。",
            "utterances": [
                {
                    "additions": {
                        "speaker": "1"
                    },
                    "end_time": 2320,
                    "start_time": 640,
                    "text": "参观八达岭长城。"
                }
            ],
            "segments": []
        },
        "metrics": [
            {
                "name": "DURATION_MS",
                "value": 3575
            }
        ],
        "token_usages": [
            {
                "task_info": "refine_chunk_001",
                "model_name": "doubao-seed-2-0-lite-260428",
                "token_usage": {
                    "prompt_tokens": 1200,
                    "completion_tokens": 80,
                    "total_tokens": 1280,
                    "cached_tokens": 0,
                    "audio_tokens": 64,
                    "audio_cached_tokens": 0,
                    "reasoning_tokens": 30
                }
            }
        ],
        "refine_info": {
            "mode": "on",
            "enabled": true,
            "status": "SUCCESS",
            "model_name": "doubao-seed-2-0-lite-260428",
            "strategy": "utterance_dynamic_chunk",
            "chunk_count": 1,
            "warnings": []
        },
        "resource": "seedasr"
    }
}

Error code #