Audio and speech

Dialogue, TTS, voice, cleanup, separation, translation and localization

• articleWhat do split edits do for continuity and pace, and when do they fail, according to open sources?
  Corrected 2026-10-24 - see C3-B1093-01; the original statements below are kept for history and are superseded where the correction says so. Partially verified: the curator re-read none of the cited sources; S6, S5, S3 rest on the producer's pinned reading. Barry Salt's counts of dialogue cuts in 35 films (from 1934 to 2014, his table) and an open Ohio State textbook agree that sound is rarely cut exactly with picture; the textbook says overlap feels natural and its absence can read as deliberately artificial. Salt says L-edits show the listener's reaction and let the pause be shorter, so the story moves faster, and that J-edits can confuse who is speaking in crowded scenes. These are one scholar's counts of within-scene dialogue cuts and a textbook's readings of clips; neither measures audience response, and neither gives a scene-transition-specific test.
• articleIf the user lives in the EEA, Switzerland or UK, which ElevenLabs Terms of Service apply, and do the voice-model and consent clauses differ from the non-EEA Terms?
  Corrected 2026-10-24 - see C3-B1128-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B678-04; the original statements below are kept for history and are superseded where the correction says so. ElevenLabs publishes two Terms of Service, both 'Last Updated: 09 October 2026': one for residents of the EEA, Switzerland or UK and one for everyone else. Section 4 (Content and User Voice Models) has 27 text blocks (heading and paragraphs) in each; after removing bullet glyphs 26 are word-for-word equal, including the 'voice you are authorized to share with us' wording, the content/voice-model licences and the training opt-out. The one differing paragraph is 4(h): non-EEA bars HIPAA protected health information, the EEA version bars Article 9 GDPR special-category data including biometric data for unique identification. The rest of the differences (withdrawal right, consumer liability, no arbitration, consu...
• articleHow does one rakugo performer make many characters distinguishable (voice, manner, kamishimo head-turns), what is documented about speech-synthesis attempts, and what survives in an audio-only or animated medium?
  Corrected 2026-10-24 - see C3-B1046-01; the original statements below are kept for history and are superseded where the correction says so. Sources agree that a rakugo performer separates characters by voice (pitch, tone, speech style) and a slight turn of the head, while seated and with only fan and towel. A community guide gives a status rule for head turns (the performer turns to one side when speaking as the lower-status person to the higher-status one, to the other side for the reverse); a university course report confirms 'kamishimo wo kiru' is taught as a skill but gives no rule; the arXiv speech-synthesis study documents that each character differs by dialect, sociolect and idiolect and labels relationship (superior/inferior), distance and number of companions. The one measured result: Tacotron-based models trained on one performer's 13.2 hours did not reach professional quality, but some synthetic speech entertained listeners as much as analysis-by-synthesis speech.
• articleWhat voice controls does Google document for Chirp 3: HD voices (pace, pause tags, SSML subset, punctuation) and where do they not apply?
  Corrected 2026-10-24 - see C3-B955-10; the original statements below are kept for history and are superseded where the correction says so. Problem: control timing of Chirp 3 HD speech. Conditions: Google Cloud Text-to-Speech Chirp 3: HD voices. Method: pace via speaking_rate from 0.25 (very slow) to 2.0 (very fast), 1.0 unadjusted. Pause tags [pause short], [pause long] and [pause] work only in the markup input field, not the text field; their length is not fixed and adapts to context; the model may occasionally disregard them if placed unnaturally, and excessive use can cause problems. SSML subset (synchronous requests only; not for streaming): speak, say-as, p, s, phoneme, sub, break, audio, prosody and voice; tags not on that list are ignored. Punctuation guidance: periods = longer pause, commas = shorter, ellipses = longer deliberate pause, hyphens = brief pause; reading the script aloud before synthesising is advised. Use (inference): for exact timing use SSML break or post-edi...
• articleWhich caption file formats does YouTube accept, and which support styling/positioning?
  Corrected 2026-10-24 - see C3-B1064-01; the original statements below are kept for history and are superseded where the correction says so. YouTube Help (captured 2026-10-23): basic formats recommended for beginners are SubRip (.srt) and SubViewer (.sbv/.sub), timing-only and editable in plain text. Advanced formats: SAMI (.smi/.sami) and RealText (.rt) with simple markup and no positioning; WebVTT (.vtt) "in initial implementation" with positioning but styling limited to b/i/u; TTML (.ttml, .dfxp; iTT subset) partially implemented; Scenarist (.scc) is YouTube's preferred broadcast format as exact CEA-608. TikTok/Instagram caption documentation not yet retrieved: gap.
• articleOpen video-to-audio (AI foley) models differ mainly by licence: MMAudio weights are non-commercial, HunyuanVideo-Foley excludes the EU/UK/South Korea, ThinkSound is listed Apache-2.0 - check before adding sound to client work
  Corrected 2026-10-24 - see C3-B1037-05; the original statements below are kept for history and are superseded where the correction says so. For adding synchronized sound effects to silent generated clips (many video models still output silent video or weak foley), three widely used open V2A models carry very different rights: MMAudio - code MIT, but "the checkpoints are licensed under CC BY-NC 4.0 and can be used for non-commercial purposes only" (R1); HunyuanVideo-Foley - Tencent Hunyuan Community License, which "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA" (R2; same family of terms as other Hunyuan releases, C3-B460-01); ThinkSound - Hugging Face metadata lists apache-2.0 (R3; GitHub licence field unset, so confirm the LICENSE file). Research direction (R4-R6): WorldSonus adds real-time (real-time factor 0.41), prompt-steerable, stereo spatial audio to interactive world models via streaming causal diffusion - relevant to the silent outputs of world models ...
• articleDo the Netflix Live Action and Animation dialogue-recording pages, the 5.1 Mix Room page and the Secondary Audio page give a dubbing studio any performance or voice direction (direction notes, emotion, effort, casting or voice match)?
  Corrected 2026-10-24 - see C3-B1051-02; the original statements below are kept for history and are superseded where the correction says so. No. In the decoded text of the four assigned pages (Live Action, Animation counterpart, 5.1 Mix Room, Secondary Audio) there is no direction-note, emotion, casting or voice-match language; they are technical capture and delivery pages. The nearest statements are indirect: Live Action says the original tonality need not be recreated and the target is a well-recorded on-set sound; both recording pages require one recording chain for characters who converse (differences can get content rejected); Animation says to commit one best take, deliver pickups on named tracks, and remove extraneous movement, breath and mouth noises. The only performance-mode rule found is on the archived delivery-spec page (version 5.1, 2024): lip-sync or voice-over dubbing as ordered by the Netflix representative, with lectern-style dubbing limited to Polish or Vietnamese.
• articleI am new to dynamics: what do threshold, ratio, attack, release, knee and make-up gain each do in a compressor, and what concrete numbers do open manuals give as worked examples or starting points?
  Corrected 2026-10-24 - see C3-B1061-01; the original statements below are kept for history and are superseded where the correction says so. Four open manuals (Audacity development manual, Ardour ACE Compressor, REAPER ReaComp, OBS) define the same core controls: threshold is where gain reduction starts, ratio is how much of the excess survives (Audacity: -10 dB threshold and 10:1 turns 0 dB into -9 dB; Ardour: 5:1 turns 5 dB into 1 dB), knee softens the transition, attack/release set time behaviour, make-up gain restores level. Numeric starting values exist only as tool defaults or streaming examples (OBS mic defaults 10:1, -18 dB, 6 ms, 60 ms); none of these four pages gives a film-dialogue setting, so an agent should treat them as definitions and arithmetic, not as a dialogue recipe.
• articleWhat can WebVTT, IMSC and YouTube caption files express, and do the specs enforce reading speed or line limits?
  Problem. Assuming a file format validates caption quality. Conditions. Choosing a delivery file. Method. WebVTT (W3C CRD, 20 May 2026): cues have start and end times, may overlap, voice spans (<v>) can name speakers and positioning exists; IMSC 1.2 (W3C Rec, 4 Aug 2020): TTML profile with ttm:agent speaker metadata and ebutts:linePadding for line-box padding. I searched both pages for 'characters per second', 'reading speed', 'words per minute', 'line length' and 'maximum number of lines': IMSC1.2 has no match and WebVTT has only an example comment on balanced wrapping, so neither specifies numeric quality limits. YouTube supports many file types and recommends .srt or .sbv as basic types, noting other types carry position and style info. Failure modes. A valid WebVTT or IMSC file can still break every limit above; Netflix additionally constrains TTML technical values (percentages, fontSize 100%). Limitations. Search was of the fetched text only; a limit could be in a linked profile...
• articleHow should speakers and non-speech sounds be labelled in captions, and how does that map from diarization output?
  Problem. Diarization gives anonymous speaker ids and no sound events. Conditions. SDH or captions for viewers who cannot hear audio. Method. Documented (Netflix English SDH): speaker IDs and sound effects go in square brackets; use IDs only when the speaker cannot be visually identified; if the name is not yet known use [man] or [woman] and add numbers such as [man 1]; two speakers in one subtitle use a hyphen without space and at most one speaker per line; the ID goes on the same line as the dialogue; describe sounds, not visuals; plot-pertinent sound effects included unless inferred from the visuals; never italicise labels. Music: lyrics between music notes, unknown song titles omitted and music described. BBC: colour is the preferred way to separate speakers, dashes are legacy and used with white text, single quotes mark out-of-vision speakers, and ttm:agent metadata may carry speaker information. Failure modes. Labelling every sound clutters the 20 cps budget; a wrong name given...
• articleShould captions follow shot changes, and what numbers do Netflix and BBC give?
  Problem. A caption that straddles a cut looks like a flash or lingers over a new shot. Conditions. Edited video where shot boundaries are known (an EDL, or shot detection run on the picture). Method. Documented: BBC says avoid subtitles that straddle a shot change by splitting the sentence or delaying the next one, merge subtitles for short shots if the merged event then ends at the second shot change, and authoring tools may use automated shot detection. Netflix: if dialogue starts on or within half a second after a cut, set the in-time to the first frame of the shot; if an out-time is within half a second before a cut, extend it to the cut keeping a 2-frame gap; if dialogue ends before a cut and no subtitle follows, still end 2 frames before the cut; as a last resort the out-time may be 12 frames after the cut provided no lingering effect. Failure modes. Snapping to cuts can shorten a caption below minimum duration or push reading speed over the cps limit; BBC lists this as a reas...
• articleHow do I turn word-level ASR or forced-alignment timestamps into caption events that meet Netflix and BBC timing rules?
  Problem. ASR returns words with start/end times, not caption events. Conditions. Word timestamps available (see C3-B636-12, C3-B844). Method. Inference-based pipeline anchored on documented rules: (1) group words into sentences/clauses by punctuation and long pauses; BBC prefers a single sentence per subtitle and clause-boundary splits for long ones. (2) Fill each event up to the line limit (for example 2 x 42 characters) but cut it earlier if the reading speed would exceed the cps limit (20 cps adult Netflix, 17 children) or if the event would exceed 7 s. (3) Set in-time at the first word start: Netflix puts it on the first frame of audio, 1-2 frames tolerance. (4) Set out-time to the last word end plus about half a second where no subtitle follows (Netflix); BBC allows up to 1.5 s hang-over. (5) Chain: where the gap to the next event is under half a second, extend the previous out-time to 2 frames before the next in-time (Netflix 'chaining'). (6) Enforce minimum 5/6 s by extending...
• articleHow should I break caption lines and split long sentences across events?
  Problem. Auto line wrapping breaks mid-phrase and hurts comprehension. Conditions. Any caption or subtitle with more than one line or event per sentence. Method. Documented: break after punctuation, before conjunctions and before prepositions (Netflix); do not separate article from noun, adjective from noun, first from last name, subject pronoun from verb, prepositional verb from preposition, verb from auxiliary/reflexive/negation. Prefer a bottom-heavy pyramid, avoid one or two words alone on the top line (Netflix English). BBC: ideal break at punctuation; each subtitle should be a single sentence, but a long sentence may extend across several subtitles, segmented at clause boundaries in preference to phrase boundaries; BBC also says well-edited text and timing matter more than perfect line breaks. Netflix: when one sentence continues across continuous subtitles, no ellipsis or dash is used if the pause is under 2 seconds. Example. BBC's own long-sentence example breaks after 'When...
• articleWhich numeric caption limits (cps, words per minute, characters per line, lines, min/max duration, gap) do Netflix, BBC and the FCC state, and how do I pick by context?
  Problem. Sources give different numbers and no single 'best'. Conditions. Choose the profile by delivery context. Method. Documented figures (page dates in bibliography): NETFLIX (streaming, English): 42 characters per line, max 2 lines, reading speed up to 20 cps adult / 17 cps children, min duration 5/6 s (20 frames at 24 fps), max 7 s, minimum 2-frame gap (all frame rates). BBC (UK broadcast and online): 160-180 words per minute (0.33-0.375 s/word), target minimum about 0.3 s per word, 37 characters per line for broadcast Teletext, online lines limited to 68% of 16:9 width (90% for 4:3, 1:1 and 9:16), max 2 lines landscape or square, 3 lines recommended for 9:16, a deliberate gap must be at least 1 s (preferably 1.5 s). FCC 47 CFR 79.1(j): states quality criteria (accuracy, synchronicity, completeness, placement) with no cps, line or duration numbers; only a live-caption accuracy metric formula (correct words / total words). Example. A 6-word line needs about 1.8-2.0 s at BBC rat...
• articleHow do word timestamps and speaker labels map into WebVTT (cue timing, in-cue word timestamps, voice spans, line breaks), and what makes a file invalid?
  Problem: turning an aligned, diarized word table into subtitle files that parse. Conditions: WebVTT per the W3C Candidate Recommendation Draft of 20 May 2026. Method: one cue block per readable chunk; cue start must be greater than or equal to every earlier cue's start and end must be greater than start; cues may overlap, which suits simultaneous speakers; speaker names go in voice spans (<v Name>), which need no closing tag when they cover the whole cue; word-level timing inside a cue uses cue timestamps, each greater than any previous one in the cue, greater than the cue start and less than the cue end; the spec's own guidance is to write cues on one line except where a break is definitely needed, since user agents insert wrapping breaks. WhisperX's CLI can also write word-highlighted SRT. Failure modes: Inference: word times from an aligner that equal the cue start or end, or that decrease, produce invalid timestamps; clamp to the open interval and drop duplicates; a mislabeled s...
• articleWhere should a narration or music stem go so it stays head-locked in a YouTube 360 upload, and why can't I ship a separate stereo track?
  Problem. Narration or music should not rotate with the viewer's head, but YouTube accepts one audio track. Conditions. YouTube Help spatial audio page with ?hl=en. Method. Mix tracks to be spatialised through the spatializer, keep head-locked tracks as stereo, export one 6-channel file laid out W, Y, Z, X, L, R at 48 kHz (PCM in MOV, or Opus in MP4 at minimum 768 kbps with mapping family 2), then run the metadata tool or equivalent. Failure modes. A separate stereo track alongside the spatial track is unsupported. Limitations. The page's routing sentence is ambiguous (see inference) and it describes the Facebook-era spatial workstation workflow. Version/date. Page copyright 2026, fetched 2026-10-24.
• articleIs there an open-licence dereverberation tool with a paper reference I can cite, and what does it need?
  Problem: Reduce room reverb on recorded speech with open code. Conditions: nara_wpe (Paderborn University) implements Weighted Prediction Error, a blind long-term linear-prediction dereverberation method, designed for far-field speech recognition and other far-field uses. Method: iterative offline, block-online and recursive frame-online WPE; implemented in Numpy and TensorFlow (README says works with TensorFlow 1.12.0); tested with Python 3.7-3.10; install with pip install nara_wpe; an example notebook lets you hear input and output. Licence: MIT, Copyright 2018 Paderborn University. Cite Drude et al., ITG 2018. Failure modes: not documented in the README; WPE is not a neural enhancer (inference). Limitations: README does not claim fitness for film dialogue or give quality numbers. Version/date: master branch README fetched 2026-10-24.
• articleWhat is the difference between a noise gate and a downward expander in FFmpeg terms, and does adrc offer both?
  Problem: Choose between hard gating and gentle expansion for dialogue. Conditions: FFmpeg agate states that range 0 turns it into an expander, so one filter spans both. Method: adrc (spectral dynamic range controller; attack default 50 ms range 1-1000, release default 100 ms range 5-2000) takes a transfer expression in dB per frequency power and ships both examples: spectral noise gate at -85 dB with attack 1 and release 5 ms, and spectral expansion at -10 dB threshold 1:2 ratio with attack 50 and release 100. Conclusion (inference): expansion reduces low levels proportionally and gives a gentler transition; a gate cuts hard, so it needs longer release for speech tails. Failure modes: the gate example with 1 ms attack is the fast, chatter-prone setting (inference). Limitations: adrc works per frequency bin, not whole-signal level. Version/date: ffmpeg.org page fetched 2026-10-24.
• articleHow do I make a noise gate or downward expander with FFmpeg's compand points, and which documented examples apply to speech?
  Problem: Gate or expand dialogue in a single filter with a transfer curve. Conditions: points are in dB relative to full scale as x/y pairs in strictly increasing x; 0/0 is assumed unless overridden. Method: attacks and decays (seconds per channel) set envelope averaging; the doc advises attack shorter than decay, typical 0.3 s and 0.8 s; soft-knee defaults to 0.01 dB; gain and volume (initial level, e.g. -90 for quiet start) are optional. Documented examples: a noise gate when noise is below the signal: compand=.1|.1:.2|.2:-900/-900|-50.1/-900|-50/-50:.01:0:-90:.1 ; a hard noise gate at -35 dB: compand=attacks=0:points=-80/-115|-35.1/-80|-35/-35|20/20 ; and an Expander curve. A curve that drops below a knee point steeper than 1:1 is a downward expander; an infinite drop is a gate (inference). Failure modes: very steep curves with attacks=0 give instantaneous gain change, audible as chatter or clicks (inference). Limitations: examples are generic audio, not tuned for speech. Version...
• articleHow does FFmpeg's agate work for dialogue (threshold, ratio, range, attack, release), and what does range=0 or mode=upward do?
  Problem: Remove low-level noise between spoken phrases without chopping words. Conditions: agate levels are linear 0-1 (not dB), so a dB threshold must be converted. Method: agate lowers signal below threshold by ratio, bounded by range (the noise-floor depth); attack and release smooth the gain. Documented defaults: range 0.06125 (allowed 0-1), threshold 0.125, ratio 2 (1-9000), attack 20 ms (0.01-9000), release 250 ms, knee 2.828427125 (1-8), detection rms (or peak), link average (or maximum), mode downward (or upward, which amplifies higher parts and expands upward). Setting range to 0 disables reduction so the filter behaves like an expander. Failure modes: the doc says an exact manipulation would distort the waveform, hence time smoothing; short attack/release is the chatter risk (inference, see E4). Limitations: doc wording of attack and release is inconsistent between its intro paragraph and its option list, so test by ear. Version/date: current ffmpeg.org page fetched 2026-1...
• articleHow do I write one performance direction that can be converted to ElevenLabs tags, SSML engines and Polly without errors?
  Problem. One script, several TTS back ends. Conditions. Providers as documented in this round. Method (inference from the sources). Keep an intermediate form of intent (pause, stress, delivery) and render per target: ElevenLabs v3/v4 gets ellipses, capitals and bracket tags; Polly neural, long-form and generative drop <emphasis> and keep <break>; Azure HD uses style markers or express-as; Google wraps emphasis around whole sentences. Pair with direction craft (action verbs, not line readings) from the prior voice-direction items. Failure modes. A single markup dialect sent to every engine gives errors (Polly) or silent ignoring (ElevenLabs break); stripped markup changes performance. Limitations. Mapping is the worker's inference; no engine output was compared. Version and date. Sources fetched 2026-10-24.
• articleWhat does Google Cloud Text-to-Speech advise about SSML emphasis and break placement?
  Problem. Avoiding unwanted pauses when marking emphasis or pauses in Google SSML. Conditions. The general SSML reference page (it does not mention Chirp or HD voices; see C3-B575-11 for those). Method. Use <emphasis level=...> (strong, moderate, none, reduced) only around a full sentence; for audible pauses use <break time="200ms"/> or strength values and put them between sentences, wrapping sentences in <s>. Failure modes. Enclosing individual words in emphasis may cause unwanted pauses. Limitations. Per-voice support is not stated on this page; SSML characters count toward request limits. Version and date. Page fetched 2026-10-24.
• articleWhich Azure voices accept SSML emphasis, and how do Azure HD voices take style and paralinguistic direction instead?
  Problem. Directing performance on Azure AI Speech. Conditions. Word-level <emphasis> and HD (Dragon HD, HD Omni, HD Flash) voices. Method. <emphasis level=...> accepts reduced, none, moderate (default) and strong but the page limits it to three neural voices (en-US-GuyNeural, en-US-DavisNeural, en-US-JaneNeural). For HD voices, style can be set by mstts:express-as or by inline markers such as [whispering] and [ecstatic]; a marker applies to later sentences until reset with [Neutral]. Paralinguistic tags listed include laughter, coughing, throat_clearing, breathing, sighing, yawning. Failure modes. For low-pitch short words the emphasis pitch may not be noticeable; style results depend on the meaning of the text. Limitations. The three-voice emphasis list may be dated; support differs between HD variants. Version and date. Page fetched 2026-10-24.
• articleDoes Amazon Polly support SSML emphasis and break on neural, long-form and generative voices?
  Problem. Choosing markup that will not error on a Polly voice type. Conditions. Standard, neural, long-form and generative voices. Method. Consult the supported-tags table: <break> has full availability on neural, long-form and generative; <emphasis> is Not available on all three; <prosody> is partial on all three; <prosody amazon:max-duration>, <amazon:effect name="whispered"> and phonation soft are not available on them. Failure modes. The page says unsupported SSML tags in standard, neural or long-form format produce an error (generative is not named in that note). Limitations. 'Partial availability' is not itemised on this page. Version and date. Page fetched 2026-10-24.
• articleHow do the SSML 1.1 break strength and time attributes interact, and what happens with neither?
  Problem. Inserting a pause or suppressing an automatic one. Conditions. SSML 1.1 processors. Method. <break/> is empty. strength takes none, x-weak, weak, medium (default), strong, x-strong; time takes CSS time values such as 250ms or 3s. With neither attribute, the processor produces a break stronger than it would otherwise have used. With both, the processor inserts the pause of the given time and takes other prosodic changes from strength. strength="none" can prevent a break the processor would add. Failure modes. The processor MAY insert a pause for strength values, so strong is not a guaranteed silence. Limitations. No maximum duration is set by the spec; engines impose their own (not checked here). Version and date. SSML 1.1 Recommendation 2010, section 3.2.3.
• articleWhat does W3C SSML 1.1 say the emphasis element must do, and which levels exist?
  Problem. Whether <emphasis> gives portable, predictable stress. Conditions. Any SSML 1.1 processor. Method. Wrap the word or phrase in <emphasis level=...>; levels are strong, moderate (default), none and reduced; none, moderate and strong are monotonically non-decreasing. 'reduced' is the opposite of emphasis (the spec's example is 'going to' becoming 'gonna'); 'none' stops a processor from stressing a word it would normally stress. The element may contain text and audio, break, emphasis, lang, lookup, mark, phoneme, prosody, say-as, sub, token, voice and w. Failure modes. The spec leaves rendering to the processor, so the same markup differs across engines and voices. Limitations. No acoustic values (Hz, dB, ms) are defined. Version and date. SSML 1.1, W3C Recommendation, 7 Sept 2010 (section 3.2.2).
• articleHow should pauses and emphasis be written for Eleven v3/v4 if SSML break is ignored, and what rules does ElevenLabs' own tag-enhancement prompt follow?
  Problem. Controlling pacing and stress on models that do not accept SSML break. Conditions. Eleven v3 and v4. Method. Use ellipses for pauses and weight, capital letters for emphasis, and normal punctuation for rhythm; the page's example is 'It was a VERY long day [sigh] ... nobody listens anymore.' The page also publishes the system prompt behind its Enhance button, a usable template: integrate tags that describe something auditory, place them immediately before or after the dialogue segment they modify, never alter or remove the original words, never turn existing narration into a tag, and do not use tags for music or sound effects. Failure modes. Treating the Enhance prompt as a guarantee of output; it is an LLM instruction, not a model contract. Limitations. The page does not quantify how long an ellipsis pause lasts. Version and date. Page fetched 2026-10-24.
• articleFor RSS-delivered Apple Podcasts episodes, should I use AAC or MP3, which bit rates by sample rate and channels, and MP4 vs ADTS?
  Problem. Choosing codec and bit rate for RSS episodes trades quality against file size and streaming cost. Conditions. RSS feed audio (distinct from Connect subscriber uploads). Method. Apple accepts MP3 or AAC, says AAC is better at the same bit rate and strongly recommends it; with AAC prefer the MP4 container over ADTS. Recommended ranges for both codecs: mono 40-80 kbps at 22.05/24 kHz and 64-128 kbps at 44.1/48 kHz; stereo 80-160 kbps at 22.05/24 kHz and 128-256 kbps at 44.1/48 kHz. If bit rate is constrained, lower the sample rate (e.g. 48 to 24 kHz), go stereo to mono, or both. Loudness conditioning must happen before encoding. Failure modes. Too low a rate gives artefacts that cannot be removed at playback; too high wastes bandwidth. Limitations. Ranges are recommendations, not stated rejection thresholds; no AAC profile is named. Version/date. Page fetched 2026-10-24T01:01:26Z.
• articleWhich file formats and channel layouts does Apple Podcasts Connect accept for subscriber audio uploads (WAV, FLAC, MP3), and when is mono refused?
  Problem. Subscriber audio uploads to Apple Podcasts Connect are rejected or degraded when container, sample rate or channel layout is wrong. Conditions. Subscriber audio uploaded in Apple Podcasts Connect, not RSS-hosted audio. Method. Use WAV, FLAC or MP3. WAV and FLAC minimum: 44.1 kHz, 16- or 24-bit; recommended: 44.1, 48, 88.2, 96, 176.4 or 192 kHz at 24-bit. Use the stereo source if one exists; if only mono exists, send two identical L/R channels because single-channel WAV/FLAC is not accepted. MP3 accepts mono or stereo: minimum 44.1 kHz at 32 kbps mono or 64 kbps stereo; recommended 44.1/48 kHz at 96-128 kbps mono or 128-256 kbps stereo. Failure modes. Mono WAV/FLAC is not accepted; the MP3 minimum is a floor, not the recommendation. Limitations. The page does not say what happens above the recommended range; loudness is covered by a prior finding. Version/date. Page fetched 2026-10-24T01:01:26Z (footer copyright 2026).
• articleWhich sources can Unreal Take Recorder capture (actors, Live Link, microphone), and what must be true for a property to be recordable?
  Problem. Planning a capture session needs to know what Take Recorder can and cannot record. Conditions. Take Recorder plugin enabled (restart needed); Live Link plugin optional for mocap. Method. Window > Cinematics > Take Recorder; +Source adds Any Actor, Camera Cuts, Level Sequences, Live Link, microphones, nearby spawned actors, Player actor or World states; or drag an actor from the Outliner. Actor sources record as Spawnable or Possessable (project default, per-source override). Microphone source writes a sound asset named with {take}/{slate} tokens, with Audio Gain in dB. Failure modes. A property that cannot be keyframed cannot be recorded; spawnable recordings of unattached actors give global-space transforms. Limitations. Documentation page only; I ran nothing in the editor. Version/date. Current dev.epicgames.com page as fetched, version not stated in extract.
• articleHow do I supply my own transcript to Apple Podcasts, what does Apple generate itself, and which languages, formats and limits apply?
  Problem: control whether Apple Podcasts shows auto-generated or publisher-provided transcripts. Conditions: episodes distributed to the Apple Podcasts catalogue (private feeds are not processed, so show no transcripts); iOS 17.4 or later; supported languages English, Danish, Dutch, Finnish, French, German, Italian, Norwegian, Portuguese, Spanish and Swedish. Method: by default Apple generates transcripts after publication and labels them Automatically generated; to use your own, change the show's transcript setting in Apple Podcasts Connect (Availability tab), link the file with the RSS transcript tag, and set episodes to Display transcripts I provide; the app then shows Provided by the show name. Accepted file types are VTT or SRT, and a VTT lets every speaker be named per line. For subscriber episodes upload VTT or SRT in Connect. Apple lets you download its transcript, edit and relink it. Best practice: correct language, spelling of host and guest names in descriptions, correct e...
• articleWhat does the Podcasting 2.0 JSON chapters file contain, which fields are required, and how do I add silent chapters, art and links?
  Problem: add navigable chapters to an episode without editing the audio file. Conditions: feed uses the podcast:chapters tag. Method: the item gets one podcast:chapters tag with url and type (JSON preferred, application/json+chapters). The file is JSON with required version and chapters array; each chapter requires only startTime (seconds, float) and may carry title, img (chapter art), url, toc, endTime and a location object (name and geo as an RFC 5870 geo URI). toc set to false makes a silent chapter that does not show in the table of contents, useful for changing artwork. Chapters are read as ascending by startTime. The file should be served with content type application/json+chapters. Benefits stated by the spec: chapters can be edited after publishing and images are fetched on playback rather than embedded. Example: a first chapter at startTime 0 titled Intro. Failure modes: out-of-order startTime values break the assumed ascending order. Limitations: the page examples show ver...
• articleWhich transcript file format should a podcast publish (WebVTT, SRT, JSON or HTML), how are speakers and word timing carried, and what does the spec recommend if I can support only one?
  Problem: choose transcript formats that work across apps and browsers. Conditions: publishing via the podcast:transcript tag. Method: the namespace transcript document names four formats. WebVTT is listed as recommended: it carries medium-fidelity timestamps, optional speaker names through a voice span tag at the start of a caption when the speaker changes, is supported natively by browsers via the track element, and is what Apple Podcasts ingests; if only one format can be supported, the document says to choose WebVTT. A maximum line length of 65 characters is recommended for caption display (the document suggests whisper-cpp options --split-on-word --max-len 65 to achieve this; I did not run them). HTML is the low-fidelity option for transcripts with little or no timing: cite for speaker, time for start of a monologue, p for content. JSON (listed as in use) has version and segments with speaker, startTime, endTime and body, and at the finest grain enables word-by-word highlighting...
• articleHow do I declare a transcript or caption file in a podcast RSS item: which attributes are required, what mime types, and what does rel=captions change?
  Problem: a podcast episode needs a machine-discoverable link to its transcript or captions. Conditions: RSS 2.0 feed using the podcast namespace (https://podcastindex.org/namespace/1.0). Method: add one podcast:transcript child per format inside the item; url and type are required; type is a mime type such as text/plain, text/html, text/vtt, application/json or application/x-subrip; language is optional and defaults to the feed's RSS language element; rel="captions" marks the file as closed captions, so timecodes are assumed present whatever the mime type. Several tags may be given for several formats. Example from the spec: a VTT file with type text/vtt, and an SRT file with type application/x-subrip and rel=captions. Failure modes: omitting language on a non-feed-language file makes it be treated as the feed language; the spec says hyperlink URLs must be https, not http. Limitations: the tag only links the file; format content rules are in the separate transcript-file document. Ve...
• articleIs audio description needed when dialogue already carries the visual information, and how do WCAG 1.2.3, 1.2.5 and 1.2.8 layer?
  Problem. Deciding whether a video needs description and which text alternative satisfies which level. Conditions. Prerecorded synchronized media, WCAG 2.2. Method. The Understanding page notes that if all important information in the video track is already in the audio track, no additional description is necessary. Standard AD means narration added during existing pauses in dialogue. 1.2.3 (A) allows audio description or a full text alternative; 1.2.5 (AA) requires audio description; 1.2.8 (AAA) requires an extended text description, adding nothing new only if 1.2.3 was met with a text description and 1.2.5 with audio. Failure modes. Treating a talking-head video as needing nothing without checking; assuming text alternative satisfies AA. Limitations. Says nothing about script timing numbers. Version/date. WCAG 2.2 Understanding page as fetched.
• articleWhat limit does WCAG state on how much a standard, user-selectable AD track can carry compared with an extended version?
  Problem. Choosing between a standard description track (G78) and a second extended version (G8). Conditions. Prerecorded media where a separate audio track or version can be offered. Method. G78: a second, user-selectable audio track adds descriptions during pauses in dialogue and sound effects, so that new information neither obscures key audio nor is obscured by loud effects. G8 states the cost: narrators sometimes must fit a lot of information into very short pauses; extended description removes that squeeze. G78 test: check that gaps in dialogue are used to convey important visual information. Failure modes. Squeezing key description into short pauses on the standard track; or obscuring effects. Limitations. G78 gives no durations; relation to the F113 usability test is in C3-B797-01. Version/date. WCAG 2.2 technique pages as fetched.
• articleWhen gaps are too short, what does WCAG 1.2.7 extended audio description require, how is it delivered, and what does the G8 test check?
  Problem. Dense dialogue leaves no gap long enough for the description that conveys the sense of the video. Conditions. Prerecorded synchronized media at WCAG Level AAA (SC 1.2.7). Method. Extended audio description periodically freezes the presentation, plays additional description, then resumes. Because this disrupts viewers who do not need it, the Understanding page says toggles are often provided, or separate versions with and without extension. G8 describes the second-version approach: the video halts (same frame repeated) while a narrator speaks. G8 test: open the extended version and check the video halts when there is not enough space between natural dialogue. Example (Understanding): a lecture where the video is paused between problems to describe sketches and gestures. Failure modes. Using extension when standard gaps suffice; the Understanding page says it is used only when the sense would be lost and pauses are too short. Limitations. No pause-length cap, no rule on how m...
• articleWhich pauses in dialogue must an audio describer use under WCAG, and what makes a pause unusable (under about 2 seconds, far from the visual event, important non-speech audio)?
  Problem. Describers need a testable rule for when a dialogue pause must carry description and when it should be left alone. Conditions. Standard (non-extended) audio description of prerecorded synchronized media, WCAG 1.2.5 (AA). Method. WCAG failure F113 says important actions, characters, scene changes and on-screen text conveyed only visually need appropriate pauses in dialogue. It notes that not all pauses are usable: pauses that are too short (example given: under 2 seconds) or not near the visual content may be unsuitable. Where visual information coincides with music or sound effects, audio ducking (lowering the overall level) can make added narration easier to distinguish. The test has two checks: all missing important visual information is conveyed, and no remaining pause in dialogue is one where description would be appropriate. Failure modes. Ducking can mask non-spoken audio that carries much of the sense of the video; some pauses should not be filled because the non-spo...
• articleCan a signal watermark prove audio came from a speech model after it passes through a neural codec?
  Problem. Using an inaudible watermark to identify generated or edited speech. Conditions. Moshi authors tested AudioSeal on Moshi speech, resampled to 16 kHz, with pink noise (sigma 0.2) and low-bitrate neural codecs (a 16 kHz RVQGAN and their Mimi). Measured result. Mean detection score over 1000 generations was 0.9999 for watermarked unmodified audio and about 0.09 (10 s) / 0.25 (1 min) for unmarked audio; pink noise lowered the 10 s score to 0.709. Failure. The authors report the Mimi codec removes the mark so watermarked audio is indistinguishable from unmarked. Method. Authors also describe audio indexing (matching) only where the generator operator can store outputs. Limitations. One watermark method, one model family; not a claim about other marks. Version/date. arXiv 2410.00037 (2024).
• articleWhat is Chion's 'added value' and how can I use its six registers to decide what a generated sound should add to a shot?
  Problem: Generated sound is often prompted by naming the object ('rain') rather than the effect on the image. Conditions: shots where the picture alone is neutral or ambiguous. Method: the lexicon lists six registers of added value (sensory, informative, semantic, narrative, structural, expressive). Pick the register the shot needs, then prompt for that: e.g. sensory (texture, temperature, energy), informative (space, time, cause), expressive (emotional colour). Because attention is on the screen, the audience attributes the sound's contribution to the image, so a sound layer can change how a shot reads without altering pixels. Example (inference): the same locked wide shot with a dry close room tone versus a cold reverberant hum reads as a different place. Failure modes: sound that contradicts the image's information (semantic clash) is read as an error, not a nuance. Limitations: secondary source; the 'six registers' is the page's summary of Chion. Version/date: as fetched.
• articleWhy does OmniHuman-1 train with text, audio and pose conditions at decreasing ratios, and what inference settings (CFG on audio/text only, 5 motion frames) does it report for long talking-body videos?
  Problem. Audio-driven human animation models that need clean lip-sync data throw away most video; stronger conditions such as pose dominate weaker ones such as audio. Conditions. You are choosing or evaluating a one-stage Diffusion Transformer avatar model (image + audio, optional pose) and want to understand why it behaves as it does. Method (OmniHuman-1, v3 of the paper). A three-stage mixed-condition training: stage 1 text+image only, stage 2 adds audio (pose dropped), stage 3 adds pose. Principle 1: stronger-conditioned tasks may reuse weaker-conditioned data, so clips failing lip-sync filters still train the text/image tasks. Principle 2: the stronger the condition the lower its training ratio, because with pose present the model leans on pose and under-learns audio; ratios for text, audio and pose are progressively halved when present. Data: 18.7K hours of in-house human video filtered for aesthetics, quality and motion amplitude, of which 13 percent met lip-sync and pose-visi...
• articleWhat is walla, who records it, and what goes wrong with generated crowd voices?
  Problem. Background crowds need voices but must not compete with the principals. Conditions. Post-production crowd scenes. Method. Walla is the murmur of a crowd, recorded by a group of actors in post (a walla or loop group). Early practice had the group repeat a nonsense word; now actors speak real, often improvised conversation tailored to the language, speech patterns and accents expected of the crowd. The UK uses 'rhubarb'. A studio guide names the main pitfall as intelligible background lines that pull focus from the main dialogue. Failure modes. Intelligible foreground-like speech; wrong language or accents for the setting. Limitations. One Wikipedia article flagged as needing citations and one generic studio guide; no primary craft source or loop group booking standards were found, and sizes of groups were not verified. AI note (inference): text-to-speech crowd tools may produce clean intelligible lines, which is the failure the guide names. Version/date. Fetched 2026-10-23.
• articleWhat do Demucs's own README and licence say: is it maintained, what licence, and what SDR does htdemucs report?
  Problem: deciding whether to depend on Demucs v4 for stem separation. Conditions: repository facebookresearch/demucs. Method: the README states the repo is no longer maintained by its author (a fork is named for bug fixes), v4 is Hybrid Transformer Demucs (default htdemucs, trained on MUSDB HQ plus 800 extra songs), 9.00 dB SDR on the MUSDB HQ test set, 9.20 dB for the sparse-attention fine-tuned model not released; htdemucs_ft is about four times slower; htdemucs_6s adds guitar and piano. Code licence MIT (Meta). Failure modes: unmaintained code may break with new torch/torchaudio versions (inference); the 6-source model has weak piano. Limitations: the licence covers code; the README does not describe training-data terms for the 800 songs or output rights. Version/date: fetched 2026-10-23.
• articleIs there one best AI dialogue cleaner for noise and reverb, according to a published blind listening test?
  Problem: choosing a noise or reverb reduction tool for production dialogue. Conditions: two Production Expert blind listening polls (noise reduction; reverb reduction), randomised processed versions, listeners voted for the most effective per clip; the reverb part used deliberately extreme examples and few clips. Method: read per-clip winners, not the aggregate. Measured result (vote shares, n of voters not stated on the page): noise test, dxRevive Pro won one clip with over 50 percent of votes, Hush Pro won broadband noise with 64.32 percent, hum was a near three-way draw (Extract: Dialogue, RX Dialogue Isolate, dxRevive Pro), roadside traffic went to Hush Pro with 38 percent; reverb test, Hush Pro led 3 of 4 clips (34.07, 53.49, 55.05 percent) and dxRevive Pro led one (40 percent vs 37.91). Failure modes: results depend on clip choice; Hush Pro came last on the hum clip. Limitations: number of listeners and clips is small, not peer reviewed, early versions of tools. Version/date: ...
• articleHow tight can machine-translated dubbing script length be to the source, and what does it cost in quality?
  Problem. Dubbing needs translation matching utterance duration. Conditions. Research on English to French, Italian, German and Spanish with a public benchmark. Method. Isometric MT trains a transformer to produce output within about plus or minus 10% of the source in character count without N-best re-ranking; a companion paper compares verbosity control methods and reports subjective tests on dubbed clips. Practical use: set a length ratio target (about 0.9-1.1 of source characters) when prompting or selecting among translations, then fit residual difference with TTS rate (C3-B792-02) and pauses. Failure modes. Length control can cost translation quality; character count is a proxy for duration. Limitations. I read only abstract pages, not numbers in the papers; results are for four language pairs. Version/date. arXiv 2112.08682 and 2110.03847. Links C3-B591-01 for the synchrony concept.
• articleHow much longer will my translated narration be than the English, and what does the W3C say about it?
  Problem. Translated voice-over often overruns the picture. Conditions. Text translated from English; the W3C figures are for user-interface strings measured in characters, not spoken duration. Method. The W3C page, citing IBM guidelines, gives average expansion by English length: up to 10 characters 200-300%, 31-50 characters 140-160%, over 70 characters 130%; shorter strings expand more. Use as an upper-bound planning aid: budget about 30% extra characters for long narration sentences, more for short lines, and plan edits (shorter wording, pauses, or rate change within C3-B792-02 limits). Failure modes. Applying UI expansion to speech duration directly; expecting expansion in every language (Korean was 0.8, Chinese 1.2 in the Flickr example). Limitations. Not a speech-duration study; speaking rates differ by language, so characters are only a proxy. Version/date. As fetched 24 Oct 2026. Links C3-B591-01 on isochrony.
• articleShould I apply SSML prosody rate to single words or to whole sentences when timing narration?
  Problem. Local rate changes inside a sentence can cause audible glitches. Conditions. Google Cloud TTS and Amazon Polly generative voices. Method. Apply prosody rate to full sentences; put intended pauses between sentences. Polly: for Generative voices prosody works only around full sentences. Google: enclosing words within a sentence may cause unwanted pauses. Combine with C3-B792-02 clamps by timing each sentence separately and scaling per sentence. Failure modes. Word-level rate tweaks producing pauses. Limitations. Standard Polly voices accept more; guidance is vendor documentation, not a measured study. Version/date. As fetched 24 Oct 2026.
• articleCan Azure TTS be told to finish in an exact duration, and what are the limits?
  Problem. Voice-over must end on a picture cut. Conditions. Azure Speech SSML with mstts:audioduration. Method. Wrap text in a voice element and add <mstts:audioduration value="20s"/>; the service speeds or slows the speech to match. Requested duration must be within 0.5 to 2 times the original unmodified audio (for 30s requested, original must be 15-60s); maximum 300 seconds; the setting applies to all text in the enclosing voice element, so split per shot. Failure modes. Out-of-range request; long multi-shot text under one value. Limitations. Documentation does not state how out-of-range is handled beyond the range rule in the part I read, and does not report quality at the ends of the range. Version/date. As fetched 24 Oct 2026. Links C3-B664-08 for cross-vendor SSML.
• articleHow far can I change TTS speaking rate with SSML prosody rate on Azure, Google and Polly before it is rejected or unreliable?
  Problem. Fitting a script into a time slot by speeding or slowing TTS needs to know accepted ranges. Conditions. SSML prosody rate (W3C), Azure Speech, Google Cloud TTS AudioConfig.speakingRate, Amazon Polly. Method. W3C: a percentage acts as a multiplier of the default rate and different voices have different natural defaults. Azure: keep rate changes within 0.5 to 2 times the original audio. Google API: speakingRate in [0.25, 2.0], and values outside return an error. Polly: rate depends on the specific voice. So compute needed ratio = natural duration / slot duration, measure natural duration per voice first, and clamp to 0.5-2 (Azure) or 0.25-2 (Google). Failure modes. Using one voice's WPM for another; extreme ratios sound unnatural (inference). Limitations. Docs state ranges not quality; no vendor states a words-per-minute figure per voice. Version/date. Pages as fetched 24 Oct 2026. Links C3-B570-08 and C3-B570-10 for SSML semantics.
• articleWhat narration speaking rate (words per minute) does the evidence support for instructional voice-over, and how strong is it?
  Problem. A narrator or TTS script needs a target words-per-minute figure. Conditions. Narrated instructional or lecture video for university students. Method. A 2026 scoping review (12 included studies out of 2679 identified) reports that cited researchers give about 150 wpm as usual speech, with comprehension spanning 150 to 270 wpm; one included experiment (Simonds et al.) found a moderate rate of 172 wpm gave higher perceived instructor credibility and nonverbal immediacy, with no significant difference in clarity or recall across rates. Use roughly 150-175 wpm as a starting script budget, then check by timing a read-through. Failure modes. Treating 172 wpm as a measured optimum (it was the moderate condition in one study); the 150-270 range is a literature claim relayed by the review, not its own measurement. Limitations. Review covers playback-speed and rate studies on university students 2017-2024, not audio description or entertainment narration; most papers did not report wo...

More