Skip to main content
Tafrigh provides extensive configuration options to optimize transcription accuracy and performance for your specific audio content.

​
Configuration overview

The transcribe function accepts an options object with three main configuration areas: 
import { init, transcribe } from 'tafrigh';

init({ apiKeys: ['your-wit-ai-key'] });

const options = {
  concurrency: 5,
  retries: 3,
  splitOptions: { /* ... */ },
  preprocessOptions: { /* ... */ },
  callbacks: { /* ... */ },
  preventCleanup: false
};

const transcript = await transcribe('audio.mp3', options);

​
Split options

Control how Tafrigh chunks your audio files for optimal transcription quality.

​
Chunk duration

The chunkDuration setting determines the maximum length of each audio chunk sent to the Wit.ai API: 
const options = {
  splitOptions: {
    chunkDuration: 60, // Split into 60-second chunks
  }
};
Actual chunk length may be shorter than chunkDuration if Tafrigh detects that splitting at the exact duration would cut through a spoken word. The library automatically finds the nearest silence to split at.
Default: 60 seconds Trade-offs:
  • Shorter chunks: More granular timestamps, faster parallel processing, but more API requests
  • Longer chunks: Fewer API requests, but less granular timestamps and longer processing time per chunk

​
Chunk minimum threshold

Filter out chunks that are too short to contain meaningful content: 
const options = {
  splitOptions: {
    chunkMinThreshold: 4, // Discard chunks shorter than 4 seconds
  }
};
Default: 0.9 seconds This prevents very short audio segments (like brief pauses or noise) from being sent to the API.

​
Silence detection

Configure how Tafrigh identifies silence for intelligent chunk splitting: 
const options = {
  splitOptions: {
    chunkDuration: 60,
    silenceDetection: {
      silenceThreshold: -30,  // Volume level considered as silence (dB)
      silenceDuration: 0.5,   // Minimum silence length to split at (seconds)
    },
  },
};

​
Silence threshold

The silenceThreshold defines the volume level (in decibels) below which audio is considered silence. Default: -25 dB Tuning guidelines:
  • Noisy background: Use a lower value (e.g., -35 dB) to only split on clear silence
  • Quiet environment: Use a higher value (e.g., -20 dB) to detect subtle pauses

​
Silence duration

The silenceDuration specifies the minimum duration of silence required to trigger a split. Default: 0.1 seconds Tuning guidelines:
  • Fast speakers with brief pauses: Use shorter duration (e.g., 0.1s)
  • Speakers with longer natural pauses: Use longer duration (e.g., 0.5s or higher)

​
Retry configuration

Configure automatic retry behavior for failed API requests: 
const options = {
  retries: 3, // Retry failed chunks up to 3 times
};
Default: 5 retries Tafrigh uses exponential backoff for retries to handle transient network failures or API rate limits gracefully.

​
Prevent cleanup

By default, Tafrigh automatically deletes temporary directories created during processing. For debugging, you can preserve these files: 
const options = {
  preventCleanup: true, // Keep temporary files after transcription
};

const transcript = await transcribe('audio.mp3', options);
// Temporary directory and chunk files remain on disk
Default: false
When preventCleanup is true, you’re responsible for manually deleting the temporary directory to avoid filling up disk space.

​
Real-world example

Here’s a complete configuration optimized for a podcast with multiple speakers: 
import { init, transcribe } from 'tafrigh';

init({ apiKeys: ['key1', 'key2', 'key3'] });

const options = {
  concurrency: 3,
  retries: 5,
  splitOptions: {
    chunkDuration: 90,  // Longer chunks for continuous speech
    chunkMinThreshold: 5,
    silenceDetection: {
      silenceThreshold: -28,  // Account for background music
      silenceDuration: 0.8,   // Split on natural conversation pauses
    },
  },
  preprocessOptions: {
    noiseReduction: {
      afftdnStart: 1,
      afftdnStop: 1.5,
      afftdn_nf: -25,
      dialogueEnhance: true,
      highpass: 200,
      lowpass: 3000,
    },
  },
};

const transcript = await transcribe('podcast-episode.mp3', options);

​
Configuration for different scenarios

1

Interview or lecture

Use longer chunks (90-120s) and longer silence duration (0.8-1.5s) to capture complete thoughts without interruption.
{
  splitOptions: {
    chunkDuration: 120,
    silenceDetection: {
      silenceThreshold: -30,
      silenceDuration: 1.5,
    },
  },
}
2

Fast-paced conversation

Use shorter chunks (30-45s) and brief silence detection (0.1-0.3s) to handle rapid exchanges.
{
  splitOptions: {
    chunkDuration: 45,
    silenceDetection: {
      silenceThreshold: -25,
      silenceDuration: 0.2,
    },
  },
}
3

Noisy environment

Adjust threshold lower and enable aggressive noise reduction.
{
  splitOptions: {
    silenceDetection: {
      silenceThreshold: -35,
      silenceDuration: 0.5,
    },
  },
  preprocessOptions: {
    noiseReduction: {
      afftdn_nf: -30,
      dialogueEnhance: true,
    },
  },
}

​
Next steps

Noise reduction

Deep dive into audio preprocessing options

Concurrency

Optimize processing speed with parallel transcription

Callbacks

Monitor progress with callback functions

Resuming failures

Handle partial failures gracefully

Build docs developers (and LLMs) love

Get started for freeTalk to us