Author: luca_admin

A practical guide based on real ADAP project rules

AI transcription tasks on Appen (CrowdGen platform) can be confusing at first, especially for new contributors.
Projects like ADAP transcription and segmentation require strict rule-based decisions, not personal interpretation.

This guide explains how these tasks really work, based on actual Appen project guidelines and real edge cases.

---

What Are Appen / CrowdGen Transcription Tasks?

On Appen’s CrowdGen platform, transcription tasks are used to train and evaluate AI speech models.

Your role is not to create a clean script for humans, but to:

• Accurately transcribe spoken content
• Identify the main speaker
• Handle interruptions and secondary speakers correctly
• Flag segmentation issues when audio is cut incorrectly

Precision matters more than readability.

---

The Main Speaker Concept (Appen ADAP Projects)

Every Appen transcription task defines a Main Speaker.

• The main speaker is the primary voice to be transcribed
• Their speech is written normally, following spelling rules
• Appen provides Main Speaker Samples to help you recognize the voice

Important detail:

Main Speaker Samples may contain more than one voice.

This is intentional:
Appen uses real audio clips so you can learn to distinguish the main speaker from background or secondary speakers.

---

Handling Other Speakers (Non-Primary Speaker)

When someone other than the main speaker talks:

• You do not transcribe their words
• You insert the special Appen tag:

<nonprimaryspeakertalking/>

This applies to:

• Short interruptions
• Background reactions (“yeah”, “mhm”)
• Comments from another person

Never guess or rewrite what the other speaker says.

---

Speech Segments vs Noise Segments on Appen

Speech Segments

A speech segment contains meaningful spoken language.

• Main speaker → transcribe text
• Other speakers → <nonprimaryspeakertalking/>

Noise Segments

Noise segments contain no meaningful speech.

Examples:

• Music
• Laughter
• Coughing
• Throat clearing
• Environmental sounds

In Appen noise segments:

• You do not transcribe words
• You use one noise tag only
• There is no main speaker

---

Segmentation Errors (Incorrect Segmentation)

Appen requires contributors to identify when audio is cut incorrectly.

You should flag Incorrect Segmentation when:

• A word is cut at the beginning or end
• A sentence is interrupted unnaturally
• The audio timing clearly does not match natural speech

Segmentation flags help Appen improve dataset quality.

---

Cut or Unclear Words

If a word is partially missing due to audio cut:

• Do not reconstruct it
• Do not guess the intended word
• Transcribe only what is audible

Appen guidelines explicitly discourage interpretation.

---

Fillers and Hesitations

If the main speaker uses fillers (“uh”, “um”, “ehm”):

• Transcribe them exactly as specified in the project rules

If a non-primary speaker produces fillers:

• Do not transcribe
• Use <nonprimaryspeakertalking/>

---

Language and Translation Rules

On Appen transcription tasks:

• Never translate
• Always transcribe in the original spoken language
• Foreign-language flags are used only when required by the project

---

Common Mistakes New Appen Contributors Make

❌ Guessing unclear words
❌ Transcribing secondary speakers
❌ Mixing speech and noise rules
❌ Ignoring segmentation problems
❌ “Cleaning up” grammar

Appen evaluates rule compliance, not writing style.

---

Final Advice for Appen Contributors

If you are unsure, ask yourself:

“What would help Appen’s AI models learn from this audio?”

Follow the rules exactly, avoid assumptions, and stay consistent.
That’s how high-quality contributors succeed on CrowdGen.

Other Tasks and Assessments

• Newsletter
• What a “Story Continuation” Assessment Looks Like (and Why It’s Used)
• Practical Guide Annotation Tasks
• How to Correctly Classify Ads Tasks

Browse Open AI Training and Data Annotation Jobs

A Practical Guide with Real Examples

When working on ad classification tasks (such as those on Appen), your goal is not to “guess,” but to answer consistently and according to the platform’s rules.

This guide explains how to approach the task step by step, avoiding the most common mistakes.

---

1️⃣ First question: Is there a recognizable advertiser?

Always start here:

👉 Is there a visible logo, name, or organization clearly shown in the ad?

• ✔️ Yes → select “Yes, there’s some advertiser info”
• ❌ No → only if there is truly no identifiable advertiser

Examples

• Apple logo → yes
• ALDI logo → yes
• NHTSA logo (government agency) → yes
• Image with no text or logo → no

---

2️⃣ Is this a store / retail chain?

This is one of the most important questions.

✔️ It is a store if:

• It’s a supermarket (ALDI, Costco)
• It’s a retail chain
• It’s a physical store

❌ It is not a store if:

• It’s a product brand (Apple, Nike)
• It’s a service company (Penske)
• It’s a government agency (NHTSA)

📌 Example

• ALDI → ✅ store
• Apple iPhone → ❌ store
• NHTSA → ❌ store

---

3️⃣ Is a brand featured?

This is where many workers make mistakes.

👉 A brand is commercial.

✔️ These are brands:

• Apple
• Nike
• Coca-Cola
• ALDI

❌ These are not brands:

• Government agencies (NHTSA)
• Public safety campaigns
• Public Service Announcements (PSAs)

📌 Simple rule

If it doesn’t sell anything, it’s probably not a brand.

---

4️⃣ Is this an advertisement for a product?

Ask yourself:
👉 Is the ad promoting something you can buy or subscribe to?

✔️ Product / service

• iPhone 16 Pro
• Equipment Protection Plan (Penske)
• Paid services

❌ Not a product

• Store logo only
• Public safety messages
• Informational campaigns

---

5️⃣ Is this a website or online service?

Select Yes only if the ad promotes:

• An app
• A website
• An online platform (Google, Facebook, etc.)

If it’s only an informational message:
👉 select “No, it’s really not.”

---

6️⃣ Is this an event or campaign?

This includes:

• Awareness campaigns
• Public safety initiatives
• PSAs (Public Service Announcements)

📌 Example

• “Don’t forget your child. Check the back seat.”
  → ✅ campaign, not a commercial event

---

7️⃣ What to write in “Primary Advertiser, Brand or Product”

Write one clear name only:

• Apple → Apple
• ALDI → ALDI
• iPhone → iPhone 16 Pro
• PSA → NHTSA

❌ No long descriptions
❌ No explanations
✔️ Just the main entity

---

⚠️ Common mistakes to avoid

• Marking a government agency as a brand
• Marking a brand as a store
• Over-analyzing company divisions (Nord/Süd, departments, etc.)
• Thinking “real-world marketing” instead of Appen rules

---

✅ Conclusion

These tasks don’t test your marketing knowledge.
They test your ability to:

• Read the ad carefully
• Follow clear classification rules
• Answer consistently

If you follow this framework, you can reduce errors to nearly zero.

Other Tasks and Assessments

• How to Work on Transcription Tasks (Appen/Crowdgen)
• Newsletter
• What a “Story Continuation” Assessment Looks Like (and Why It’s Used)
• Practical Guide Annotation Tasks

Browse Open AI Training and Data Annotation Jobs

This post is a hands-on guide to completing Cardamom HybX annotation tasks correctly, based on real examples and frequent QA feedback.
It does not replace the official guidelines, but helps you understand what to annotate, where, and why, avoiding common rejection reasons.

---

🎯 Task Objective

Each task requires two complementary actions:

1. Compare Reference vs Annotation audio
   • Voice Characteristics
   • Speaking Style
2. Add inline tags for local events
   • Details of Speech Delivery
   • Annotation audio only

---

✅ Audio Validation: when audio is VALID or INVALID

VALID audio (most cases)

Audio is VALID if:

• Only one main speaker is present
• The speaker reads, narrates, or acts characters
• The speaker changes voice to impersonate someone
• There are background noises or brief interjections
• The transcript has small mismatches (1–2 words or a sentence at start/end)

⚠️ Invalid tasks are exceptions, not the rule.

INVALID audio (rare cases)

Audio is INVALID only if:

• There is a fluid conversation between two or more people
• The transcript is a 100% mismatch with the audio

If in doubt → do NOT invalidate.

---

🎙️ Voice Characteristics (Voice Variation)

What this field is about

Voice Characteristics describe how the voice sounds physically, compared to the reference.

Ask yourself:

Does the voice itself sound different?

Examples of valid dimensions:

• Timbre (metallic, rough, soft, nasal, warm)
• Pitch (higher / lower)
• Stability (trembling, uneven)
• Articulation (more precise / more blurred)
• Emission or resonance

When to select “There is variation”

If the sound of the voice changes, even slightly:

• More metallic → more natural
• More compressed → more open
• More rough → smoother

When to select “No difference”

If the voice is the same and only style or interpretation changes.

⚠️ Do not describe:

• Acting style
• Emotion
• Rhythm
• Expressiveness

Those go elsewhere.

---

🎤 Speaking Style

What this field is about

Speaking Style describes how the message is delivered, not how the voice sounds.

Ask yourself:

How is the speaker performing or delivering the speech?

Valid aspects:

• Theatrical vs natural
• Formal vs casual
• Confident vs insecure
• Controlled vs expressive
• Narrative, dramatic, conversational

Key rules

• Always write comparatively
• Use complete sentences
• Avoid the word “tone” (very common rejection reason)
• 20–50 words is the safe range

✔ Correct:

The delivery feels more theatrical and expressive, with a less spontaneous and more constructed performance that increases dramatic emphasis.

❌ Incorrect:

The tone is more dramatic. (→ rejected)

---

🏷️ Details of Speech Delivery (Inline Tags)

What tags are for

Tags describe local, noticeable events in the Annotation audio only.

Examples:

• Emphasis on a word
• Deliberate pause
• Hesitation
• Sighs
• Syllable-by-syllable pronunciation
• Chanting / sing-song delivery

---

How to place tags

• Insert tags only in the space between words
• Place the tag immediately before the event
• Never tag words or punctuation
• Never quote the transcript text

---

Common tag examples (safe)

• enfasi marcata su una parola
• pausa deliberata
• esitazione marcata
• sospiro udibile durante l’emissione
• frase pronunciata con sospiro continuo
• scansione sillabica per enfasi
• intonazione cantilenante su una parola

✔ Example:

… ma se ride anche 🏷️ enfasi marcata sul soggetto lei!

---

🚫 Common Mistakes That Cause Rejection

• Mixing Voice Characteristics and Speaking Style
• Using the word “tone” in Speaking Style
• Writing very short descriptions
• Using bullet points instead of sentences
• Mentioning the reference or annotation audio explicitly
• Describing audio quality or recording issues
• Guessing age, gender, identity, or accent
• Forgetting to add at least one tag

---

🧠 Mental Checklist Before Submitting

• Did I describe voice sound only in Voice Characteristics?
• Did I describe delivery and performance only in Speaking Style?
• Did I add at least one inline tag?
• Are my descriptions comparative, clear, and 20–50 words long?
• Did I avoid forbidden words?

If yes → you’re good to submit ✅

---

TL;DR

• Voice Characteristics = how the voice sounds
• Speaking Style = how the speech is delivered
• Tags = local events in the annotation audio
• Invalid tasks are rare
• When unsure → annotate, don’t invalidate

Other Tasks and Assessments

• How to Correctly Classify Ads Tasks
• How to Work on Transcription Tasks (Appen/Crowdgen)
• Newsletter
• What a “Story Continuation” Assessment Looks Like (and Why It’s Used)

Browse Open AI Training and Data Annotation Jobs

Story continuation tasks are a common type of assessment used in AI training and evaluation projects.
They are not creative writing exercises in the traditional sense. Instead, they are designed to test how well you can follow constraints while maintaining language quality and coherence.

1. What the task usually looks like

You are given:

• the beginning of a story (often already rich in tone and setting)
• a randomly generated sentence
• a word limit (usually around 200 words)

Your task is to continue the story in English and integrate the random sentence naturally into the text.

The sentence is often intentionally awkward or unrelated.
That is not a mistake — it’s the core of the test.

---

2. What the assessment is actually testing

Despite appearances, this task is not about imagination. It evaluates:

• grammatical accuracy and fluency
• narrative consistency (tone, pacing, point of view)
• instruction-following
• the ability to integrate unrelated input without breaking context

In AI-related work, this mirrors how models must handle noisy or artificial constraints while still producing usable output.

---

3. How to integrate the random sentence correctly

The most common mistake is forcing the sentence into dialogue or exposition where it doesn’t belong.

A safer approach is to use it as:

• an internal thought
• a fleeting memory
• a moment of contrast or irony

This allows the sentence to exist naturally without dominating the story.

Example strategy:

• high-tension scene → brief absurd or trivial thought → return to tension

---

4. What to avoid

• Ignoring the sentence
• Quoting it without integrating it
• Changing its wording
• Overexplaining why it appears

Assessors look for control, not justification.

---

5. Final tip

Write with intention.
Every sentence should serve the story and the instructions.

In these assessments, clarity and discipline matter more than style.

Other Tasks and Assessments

• Practical Guide Annotation Tasks
• How to Correctly Classify Ads Tasks
• How to Work on Transcription Tasks (Appen/Crowdgen)
• Newsletter

Browse Open AI Training and Data Annotation Jobs