Wjs Transcribing Audio

wjs-transcribing-audio

Spoken audio in → timestamped SRT in the same language out. This skill stops at the source-language SRT. Translation to another language is the next skill (/wjs-translating-subtitles).

When to use

• User provides a video or audio file and wants a transcript / SRT in the source language.
• User already has a translated SRT and the source SRT is missing.
• User asks "做 SRT" / "make subtitles" / "出逐字稿" with no translation step requested yet.

When NOT to use

• Source-language SRT already exists → skip straight to /wjs-translating-subtitles.
• User wants the transcript in a different language than spoken → run this skill first, then /wjs-translating-subtitles.
• User wants only the dub or burn-in → if SRT exists, skip; otherwise run this first.

Routing: which engine

For Chinese, do not default to Whisper unless the user explicitly asks for it or Volcano is unavailable. This is a deliberate routing decision — see user's memory on Chinese ASR priority.

OpenAI Whisper API path (non-Chinese, and Chinese fallback)

The key principle: do not request response_format=srt. Whisper cue-segmentation fails on long monologues (30-second blob cues) and quiet stretches (loop hallucinations). Request word-level timestamps and assemble cues yourself — the post-processing is deterministic and free.

Why not response_format=srt

Two failure modes that wreck whisper-1 SRT output on long content:

1. 30-second blob cues. In long monologues, whisper-1 with response_format=srt emits one cue covering the full 30s condition_on_previous_text window. Transcript is fine; timing is unusable for on-screen reading.
2. Loop hallucination on quiet tails. Greedy temperature=0 on low-energy audio produces "你如果不把拥抱浪费写在这上面,你很难的" repeated 50 times.

Both stem from letting Whisper decide cue boundaries. Fix: word-level timestamps + your own punctuation-aware assembler.

Calling the API

# 1. Compress for upload — 64kbps mono MP3 is plenty for speech.
#    OpenAI limit is 25MB per request; chunk into 10-min pieces
#    (≈4.5MB at 64kbps) for resilience under flaky proxies.
ffmpeg -hide_banner -loglevel error -y \
  -ss <start> -t 600 -i input.mp4 \
  -vn -ac 1 -ar 16000 -c:a libmp3lame -b:a 64k chunk.mp3

# 2. Request word-level timestamps. Do NOT request response_format=srt.
import httpx, os
data = {
    "model": "whisper-1",
    "language": "es",                        # pin source language; never auto-detect
    "response_format": "verbose_json",
    "timestamp_granularities[]": "word",     # ← the critical flag
    "temperature": "0.2",                    # enable fallback chain (anti-loop)
}
with open("chunk.mp3", "rb") as f:
    r = httpx.post(
        "https://api.openai.com/v1/audio/transcriptions",
        headers={"Authorization": f"Bearer {os.environ['OPENAI_API_KEY']}"},
        data=data,
        files={"file": ("chunk.mp3", f, "audio/mpeg")},
        timeout=600.0,
    )
r.raise_for_status()
j = r.json()
words    = j["words"]      # [{"word": "hola", "start": 0.12, "end": 0.34}, ...]
segments = j["segments"]   # see surprise below

Surprise: words[] has no punctuation, segments[] is inconsistent

Whisper's words[] array typically has no punctuation in word["word"] — each entry is a bare token like "做", "个", "测", "试". Punctuation, when present, lives only in segments[] text field.

Worse, segments[] text is inconsistently punctuated across chunks of the same file: chunk 0 of a 79-min podcast might emit 285 bare segments ("做个测试" "你在" "呵呵") at 1-2s each with no punctuation; chunk 7 might emit 34 segments at 14-30s each with punctuation. Both behaviors ship in the same API response.

So the right recipe combines both: use segments[] for natural pause boundaries (already aligned to breath), but treat them as raw input to your own cue assembler, which uses word timestamps to split anywhere the segments are too long.

Cue assembly recipe

TARGET_DUR = 3.0   # try to make cues this long
MAX_CUE_DUR = 5.0  # never exceed
MAX_CHARS = 18     # ~one line at Fontsize 14 on 1080-wide vertical
MAX_GAP = 1.0      # silence threshold → force cue boundary
MIN_PIECE = 0.3    # below this, merge with neighbor
SPLIT_PUNCT = set("，。！？；,.;!?")

# Step A: merge short segments[] toward TARGET_DUR (use segments,
#         not words — Whisper's segment boundaries are already
#         pause-aligned).
def assemble(segments, offset):
    cues, buf = [], []
    def flush():
        if buf:
            cues.append((buf[0]["start"]+offset, buf[-1]["end"]+offset,
                         "".join(s["text"].strip() for s in buf)))
            buf.clear()
    for s in segments:
        dur = s["end"] - s["start"]
        # Long single segment WITH internal punct → split standalone
        if dur > MAX_CUE_DUR and any(c in s["text"] for c in SPLIT_PUNCT):
            flush(); cues.extend(split_long_segment(s, offset)); continue
        if not buf: buf.append(s); continue
        if (s["start"] - buf[-1]["end"]) >= MAX_GAP \
           or (buf[-1]["end"] - buf[0]["start"]) >= TARGET_DUR \
           or (s["end"] - buf[0]["start"]) > MAX_CUE_DUR:
            flush()
        buf.append(s)
    flush(); return cues

# Step B: final pass — split every internal comma/period to its own cue
#         (proportional timestamps by char position). Coalesce pieces
#         shorter than MIN_PIECE forward.

# Step C: any cue still > MAX_CHARS gets split at the largest inter-word
#         gap using words[] timestamps. Recursive until under cap.

Tweak TARGET_DUR and MAX_CHARS to platform reading rhythm. The 18-char cap matters for burn-in on vertical 1080×1920 at Fontsize=14 — longer wraps to multiple unreadable lines.

Operational details

• Auth: credentials live in ~/code/.env. Load with set -a; source ~/code/.env; set +a before invoking.
• SOCKS proxy on this machine: httpx needs the socksio extra — use uvx --with httpx --with socksio python ... (without it you get ImportError: Using SOCKS proxy, but the 'socksio' package is not installed).
• Chunking: 10-min pieces at 64kbps mono MP3 (~4.5MB each) are the reliability sweet spot. 20-min chunks (~9MB) sometimes RST under flaky proxies. Concurrency max_workers=2 is more reliable than 4.
• Retry: every API call wrapped in 5× exponential backoff (time.sleep(min(2**n, 30))) — RemoteProtocolError: Server disconnected is common and transient.
• Offset stitching: each chunk's words come back with timestamps relative to that chunk. When merging, add the chunk's absolute start offset to every word's start/end before assembling cues.
• Loop guard (belt + suspenders): even with temperature=0.2, occasionally a sub-chunk still loops. After assembly, run a loop-detector on each cue's text — if any phrase of length 8–40 chars repeats 3+ times consecutively, drop the cue.

Anti-patterns (do not do)

• ❌ Do not request response_format=srt for content longer than ~2 minutes.
• ❌ Do not "fix" bad cues with a second API call. If you got blob cues or loop hallucinations from your first call, redo with word-level granularity once — don't re-transcribe just the broken sub-range.
• ❌ Do not use temperature=0 on potentially-quiet audio (yoga, spiritual content, podcast outros). Greedy decoding loops. 0.2 enables the fallback chain.
• ❌ Do not skip language=.... Auto-detect occasionally swaps Chinese→Japanese or Spanish→Portuguese on the first 30 seconds and the whole transcript is then wrong.

Volcano (豆包) ASR path — preferred for Chinese

Volcano ASR routinely beats Whisper on Mandarin accuracy (recognition rate, punctuation, named entities). Use this as the default for zh-* source.

⛔ NEVER use 飞书妙记 / lark-minutes for ASR. It only gives turn-level (speaker-turn) timestamps, not per-word timing — subtitles built from it drift by seconds and break at unnatural places. The user's standing rule: "以后不要用飞书妙记来做ASR，不能作为SRT使用，记住". There is no 飞书妙记 fallback. If Volcano is unavailable, fall back to the OpenAI Whisper word-level path above (pin language=zh).

Use the STREAMING WebSocket API — pushes bytes, needs NO public URL

The file / 录音文件识别 / MediaKit APIs all require a publicly reachable HTTP(S) audio URL. The user rejected URL-hosting ("不要再用什么从服务器端去 download 的这 mp3 这样的模式" — tunnels fail on their hotspot). The 大模型流式语音识别 (bigmodel streaming) API sidesteps the URL entirely by pushing raw PCM bytes over a WebSocket. This is the working path. It returns per-word ms timestamps.

Bundled scripts (use these — they are the verified working path):

# 1. Transcribe: pushes 16k mono PCM bytes over WebSocket → ASR JSON
#    (decodes any input via ffmpeg; .pcm passes straight through)
export VOLC_ASR_APPID=…  VOLC_ASR_ACCESS_TOKEN=…   # credentials live with the user
python3 scripts/volc_asr_stream.py <clip.mp4|wav|mp3|pcm> <out.asr.json>

# 2. Build a clean, word-timed SRT from the ASR JSON
python3 scripts/build_srt_from_asr.py <out.asr.json> <out.srt> [max_chars=18]
#    Segmentation knobs (optional):
#      --max-chars N   raise to keep cues whole / break on punctuation, not mid-sentence
#      --soft-min N    min chars before a ，、； flushes a cue (default 8)
#      --strip-punct   remove ALL punctuation from displayed text (clean subtitle look)

• Endpoint: wss://openspeech.bytedance.com/api/v3/sauc/bigmodel
• Headers: X-Api-App-Key:{appid}, X-Api-Access-Key:{token}, X-Api-Resource-Id:volc.bigasr.sauc.duration, X-Api-Connect-Id:{uuid}
• Binary v3 framing: 4-byte header [(ver<<4)|hdrsize, (msgtype<<4)|flags, (ser<<4)|comp, 0]. Full-client(0x1)+POS_SEQ sends gzipped JSON config; audio(0x2) packets send gzipped PCM chunks (200ms = 6400 bytes @16k mono s16le); last packet uses NEG_WITH_SEQ(0x3) with negative seq. Server full-response(0x9) returns gzipped JSON.
• Config JSON: {user:{uid}, audio:{format:"pcm",rate:16000,bits:16,channel:1,codec:"raw"}, request:{model_name:"bigmodel", enable_punc:true, enable_itn:true, show_utterances:true}}. Do NOT set result_type:"single" — that returns only the latest sentence each frame; default mode accumulates all utterances.
• Result shape: result.utterances[], each with text (punctuated) + words[] (token + ms start_time/end_time). Latin tokens like "AI" come back with start=end=0 — build_srt_from_asr.py forward/backward-fills from neighbours.
• build_srt_from_asr.py also: splits cues on punctuation (HARD 。！？ flush; SOFT ，、； flush past --soft-min chars; hard cap --max-chars), optionally strips all punctuation from the display (--strip-punct), drops 呃/嗯/唉 fillers, and collapses immediate duplicate short tokens (才才→才). Every cue is timed by its first/last word so it sits exactly on the spoken audio — no drift. To avoid mid-sentence breaks, raise --max-chars so boundaries fall on punctuation rather than the char cap.
• Per-clip transcription: transcribe each clip's own audio (16k mono PCM) so the SRT timestamps come out clip-relative.
• Credentials: live with the user (豆包语音引擎). Env names accepted by the script: VOLC_ASR_APPID/VOLC_APPID and VOLC_ASR_ACCESS_TOKEN/VOLC_TOKEN; FFMPEG_BIN optional.
  • This user's 火山 ASR and TTS share ONE credential set — same appid + access token work for both services. They're stored in ~/code/.env as VOLC_TTS_APPID / VOLC_TTS_ACCESS_TOKEN (plus VOLC_APPID in ~/.zshrc), with VOLC_ASR_APPID / VOLC_ASR_ACCESS_TOKEN aliases appended pointing at the same values. So set -a; source ~/code/.env; set +a is enough — no separate ASR token to hunt for. If the aliases are ever missing, recreate them from the TTS values (they're the same secret).
• Dead ends (don't retry): MediaKit asr-subtitles (doc 6448/2381968) still needs a public URL and a separate MediaKit API key. The 录音文件识别 file API needs a public URL.