Music trend detection

You will build an early-warning detector for sounds that are taking off. A song moves on the creation side first: creators pick it up on TikTok, and only later does it reach the charts. Measuring how many videos use a sound, and how far those videos travel, catches the move while it is still early.

How can you catch a rising sound before it charts?

Combine two signals. GET /v1/tiktok/song returns the sound with how many videos use it. GET /v1/tiktok/song/videos returns videos that use it, with their view counts. A log-weighted mix of the two, stored per day, shows which songs are accelerating.

What you need

Resolving the id

TikTok keys sounds by id, not by name, and the id is stable per track. Look it up once and store it.

• TikTok clipId: GET /v1/tiktok/search/music (10 credits) searches sounds by keyword. Each sound comes back as a post: post.id is the clipId, post.content.text the title, and post.ext.video_count how many videos use it. sort_by=most-used puts the biggest first.

The code

# music_heat.py
# A heat score for one track. Store it per day and watch the delta.
# Run with: INSIGHTSOCIAL_API_KEY=isk_live_... python music_heat.py
import math
import os
from concurrent.futures import ThreadPoolExecutor

import requests

KEY = os.environ["INSIGHTSOCIAL_API_KEY"]
BASE = "https://api.insightsocial.app/v1"

# Resolve the id once (see above), then keep it.
TIKTOK_CLIP_ID = "7364498342501435408"   # "Espresso", Sabrina Carpenter

def get(path, **params):
    res = requests.get(f"{BASE}/{path}", params=params,
                       headers={"x-api-key": KEY}, timeout=120)
    return res.json()

def views_on_page(body):
    """Sum of views on the page, and how many rows showed a view count."""
    if not body["success"]:
        return 0, 0
    views = [i["post"]["engagement"]["views"] for i in body["data"]["items"]]
    known = [v for v in views if v is not None]  # null means hidden, not zero
    return sum(known), len(known)

with ThreadPoolExecutor(max_workers=2) as pool:
    sg = pool.submit(get, "tiktok/song", clipId=TIKTOK_CLIP_ID)
    vd = pool.submit(get, "tiktok/song/videos", clipId=TIKTOK_CLIP_ID)
    song, videos = sg.result(), vd.result()

use_count = (song["data"]["post"]["ext"] or {}).get("use_count") or 0 if song["success"] else 0
sample_views, on_page = views_on_page(videos)

# One page is a sample, not a census. Weight it with the usage count, then track the trend.
heat = 0.5 * math.log10(use_count + 1) + 0.5 * math.log10(sample_views + 1)

print({
    "use_count": use_count,
    "sample_views": sample_views, "videos_on_page": on_page,
    "heat_score": round(heat, 2),
    "credits_used": song.get("credits_used", 0) + videos.get("credits_used", 0),
})

What you get back

{'use_count': 2140000, 'sample_views': 52700000, 'videos_on_page': 30,
 'heat_score': 7.03, 'credits_used': 20}

The figures are illustrative. The video list returns one page, not every video that uses the sound, so the view sum is a reach sample. A row whose view count is hidden (null) adds nothing to the sum and is not counted in videos_on_page. Compare a track's score with its own history, not with an absolute bar.

What to change

• The id. Swap in the track you are watching. Budget 10 credits once per new track if you still need to look it up.
• Weight the usage count's change. post.ext.use_count is the cleanest creation-velocity number TikTok publishes; its day-over-day change is worth a term of its own in the score.
• See how the sound is being used. Read the captions yourself: post.content.text on each row of tiktok/song/videos. Counting rows by post.published_at day gives you a date histogram of the page.
• Find candidates to score. GET /v1/tiktok/search/music with sort_by=most-used (10) surfaces the most-used sounds for a keyword.
• Deepen the sample. Send pagination.next_cursor back as cursor, with the same id, to read more pages at 10 credits each, and adjust the weights to match.
• Track the delta, not the level. Store heat_score per track per day. A rising score is the signal. A high, flat score is a song that already broke.

Related