Index-Echo bilingual SRT from a Chinese video to Sume cues

To subtitle a Chinese-language video in English, run Index-Echo S2TT on the file, convert its SRT into caption cues, and burn the English line with Sume. Index-Echo S2TT is the speech-to-text-translation model in Bilibili's Index-Translate family. It transcribes Chinese speech and translates each sentence in one model, with per-sentence timestamps, and writes a bilingual input.srt.

The packaged interface on its model card covers Chinese into English, Japanese or Spanish. This post uses English because Sume's Latin caption styles draw it cleanly.

What does Index-Echo produce?

Per the model card, the command takes a video and a target language and writes an SRT with the Chinese transcript and the translation, plus a per-window JSONL summary. Audio is processed in windows capped at 60 seconds, and longer files are segmented in sequence. The card lists Apache-2.0 and about 10 GB of VRAM at bf16, and it needs ffmpeg on the path. It also warns that timestamp quality can vary with audio conditions and that greedy decoding may repeat on unfamiliar input.

How do you turn the SRT into Sume cues?

Sume's video captions endpoint accepts cues with text, start and end in seconds, so a bilingual SRT is a parse away. Check which line of each block is Chinese and which is English in your file before you run this; the helper assumes the translation is the last line of each block.

import re

def to_seconds(stamp):
    h, m, rest = stamp.split(":")
    s, ms = rest.split(",")
    return int(h) * 3600 + int(m) * 60 + int(s) + int(ms) / 1000

def srt_to_cues(path):
    cues = []
    text = open(path, encoding="utf-8").read().strip()
    for block in re.split(r"\n\s*\n", text):
        lines = block.splitlines()
        if len(lines) < 3:
            continue
        start, end = lines[1].split(" --> ")
        cues.append({"text": lines[-1].strip(),
                     "start": to_seconds(start.strip()),
                     "end": to_seconds(end.strip())})
    return cues

print(srt_to_cues("out_dir/input.srt")[:2])

What limits apply on the Sume side?

A caption request takes at most 200 cues, each up to 400 characters, with start and end times inside 60 seconds, and cues cannot be mixed with script_text. The 60-second ceiling matches Index-Echo's window, but a longer video needs one job per 60-second piece, with cue times shifted back to start at zero. A standalone job is priced at $0.20 for videos up to 60 seconds under the current estimate.

The video_url must be public HTTPS. Name a Latin style such as slam for English text. Then post the converted cues as in the SRT walkthrough.

What should you review by hand?

• Sentences the model repeated, a known greedy-decoding failure.
• Cues whose timestamps drift from the speech in noisy audio.
• Names and product terms, which are worth a glossary check.
• English lines that run long for their cue, which a reading-speed check will flag.