AI_TRANSCRIBE

AI_TRANSCRIBE

Overview

Transcribes audio and video files stored in a stage, extracting text, timestamps, and speaker information.

Syntax

AI_TRANSCRIBE(
  file_reference,
  [options]
)

Parameters

• file_reference (FILE): Reference to audio/video file in stage (using TO_FILE)
• options (OBJECT): Optional settings for language, speaker detection, timestamps

Use Cases

• Meeting transcription
• Call center analysis
• Video content indexing
• Podcast transcription
• Accessibility (captions)
• Compliance and documentation

Supported File Formats

• Audio: MP3, WAV, M4A, AAC, FLAC
• Video: MP4, MOV, AVI, MKV

Code Examples

Example 1: Basic Audio Transcription

SELECT 
    file_name,
    AI_TRANSCRIBE(
        TO_FILE('@audio_stage/' || file_name)
    ) AS transcription
FROM DIRECTORY('@audio_stage')
WHERE file_name LIKE '%.mp3';

Output:

{
  "text": "Welcome to today's meeting. Let's discuss the Q4 roadmap.",
  "duration": 125.5
}

Example 2: Transcription with Timestamps

SELECT 
    AI_TRANSCRIBE(
        TO_FILE('@meetings/team_meeting_2024_01.mp4'),
        {'include_timestamps': true}
    ) AS transcription;

Output:

{
  "segments": [
    {"start": 0.0, "end": 3.5, "text": "Good morning everyone"},
    {"start": 3.5, "end": 8.2, "text": "Let's review the quarterly results"},
    {"start": 8.2, "end": 15.1, "text": "Revenue increased by 23 percent"}
  ],
  "duration": 1825.3
}

Example 3: Speaker Diarization

SELECT 
    meeting_id,
    AI_TRANSCRIBE(
        TO_FILE('@recordings/' || filename),
        {
            'identify_speakers': true,
            'include_timestamps': true
        }
    ) AS transcription
FROM meeting_recordings
WHERE meeting_date >= CURRENT_DATE - 7;

Output:

{
  "segments": [
    {"speaker": "Speaker 1", "start": 0.0, "end": 5.2, "text": "Let's begin the standup"},
    {"speaker": "Speaker 2", "start": 5.5, "end": 12.3, "text": "I completed the API integration"},
    {"speaker": "Speaker 1", "start": 12.5, "end": 18.0, "text": "Great work. What's next?"}
  ]
}

Example 4: Multilingual Transcription

SELECT 
    AI_TRANSCRIBE(
        TO_FILE('@podcasts/episode_45.mp3'),
        {'language': 'es'}  -- Spanish
    ) AS transcription;

Example 5: Batch Processing Call Recordings

CREATE TABLE call_transcriptions AS
SELECT 
    call_id,
    customer_id,
    call_date,
    AI_TRANSCRIBE(
        TO_FILE('@call_recordings/' || recording_file),
        {'identify_speakers': true}
    ) AS transcription,
    CURRENT_TIMESTAMP() AS transcribed_at
FROM customer_calls
WHERE transcribed_at IS NULL
  AND call_date >= CURRENT_DATE - 30;

Data Output Examples

Customer Service Call

Input: 5-minute customer support call
Output:
{
  "duration": 305.2,
  "segments": [
    {"speaker": "Agent", "text": "Thank you for calling support. How can I help?"},
    {"speaker": "Customer", "text": "I'm having trouble logging into my account"},
    {"speaker": "Agent", "text": "I can help with that. What error message do you see?"}
  ]
}

Podcast Episode

Input: 45-minute podcast episode
Output:
{
  "duration": 2701.8,
  "text": "Welcome to Tech Talk. Today we're discussing artificial intelligence...",
  "word_count": 7250
}

Model Information

• Billing: 50 tokens per second of audio
• Languages: Multiple languages supported
• Speaker Detection: Up to 10 speakers

Limitations & Considerations

File Size and Duration

• Maximum file size varies by region
• Longer files take more time to process
• Consider splitting very long recordings

Audio Quality

• Better audio quality = better transcription
• Background noise can affect accuracy
• Clear speech improves results

Cost

• Billed at 50 tokens per second of audio
• A 10-minute file = 600 seconds = 30,000 tokens
• Speaker detection may increase cost

Processing Time

• Not real-time; designed for batch processing
• Processing time scales with file duration
• Use async patterns for large batches

Regional Availability

• AWS US West 2: ✓
• AWS US East 1: ✓
• Azure East US 2: ✓
• EU regions: ✓

Best Practices

1. Create Dedicated Stages

CREATE STAGE audio_transcriptions
  DIRECTORY = (ENABLE = true)
  ENCRYPTION = (TYPE = 'SNOWFLAKE_SSE');

PUT file://local_audio/*.mp3 @audio_transcriptions AUTO_COMPRESS = FALSE;

2. Store Results in Tables

CREATE TABLE meeting_transcripts (
    meeting_id VARCHAR,
    recording_file VARCHAR,
    transcription VARIANT,
    transcription_date TIMESTAMP DEFAULT CURRENT_TIMESTAMP()
);

INSERT INTO meeting_transcripts
SELECT 
    meeting_id,
    filename,
    AI_TRANSCRIBE(TO_FILE('@meetings/' || filename))
FROM meeting_recordings;

3. Extract and Analyze Transcriptions

-- Extract text from transcription
SELECT 
    meeting_id,
    transcription:text::STRING AS full_text,
    transcription:duration::FLOAT AS duration_seconds
FROM meeting_transcripts;

-- Analyze with other AI functions
SELECT 
    meeting_id,
    AI_SENTIMENT(transcription:text::STRING) AS meeting_sentiment,
    AI_EXTRACT(
        transcription:text::STRING,
        'Extract action items and decisions made'
    ) AS key_outcomes
FROM meeting_transcripts;

4. Process in Batches

-- Process recordings in manageable batches
CREATE OR REPLACE TASK transcribe_daily_recordings
  WAREHOUSE = compute_wh
  SCHEDULE = 'USING CRON 0 2 * * * UTC'  -- 2 AM daily
AS
  INSERT INTO transcriptions
  SELECT 
      file_url,
      AI_TRANSCRIBE(TO_FILE(file_url))
  FROM pending_transcriptions
  LIMIT 100;  -- Process 100 at a time

Common Use Cases

Meeting Minutes Generation

SELECT 
    meeting_id,
    meeting_title,
    AI_COMPLETE(
        'claude-4-sonnet',
        'Create meeting minutes from this transcript: ' || transcription:text::STRING
    ) AS meeting_minutes
FROM meeting_transcripts
WHERE meeting_date >= CURRENT_DATE - 7;

Call Center Quality Analysis

SELECT 
    agent_id,
    COUNT(*) AS calls_analyzed,
    AVG(AI_SENTIMENT(transcription:text::STRING)) AS avg_sentiment,
    AI_AGG(
        transcription:text::STRING,
        'Identify common customer complaints'
    ) AS common_issues
FROM call_transcriptions
GROUP BY agent_id;

Compliance Monitoring

SELECT 
    call_id,
    AI_FILTER(
        'Does this call mention required compliance disclosures?',
        transcription:text::STRING
    ) AS compliance_check
FROM call_transcriptions
WHERE compliance_check = false;

Related Functions

• AI_EXTRACT - Extract specific information from transcripts
• AI_SUMMARIZE_AGG - Summarize multiple transcripts
• AI_SENTIMENT - Analyze transcript sentiment
• TO_FILE - Reference files in stages