The open TTS line shipped consent terms the commercial one hasn’t
AI4Bharat’s IndicF5 update points at a gated 22-language successor while carrying explicit cloning-consent terms; EmbeddingGemma 2 takes retrieval on-device; the Decisions API prices the small fast judge; Sonnet 4.5 gets a retirement date — a two-desk week.
AI4Bharat’s IndicF5 update recommends its gated 22-language successor and carries cloning-consent terms its commercial rival still hasn’t written; EmbeddingGemma 2 puts modular multimodal retrieval on-device; OpenAI’s Decisions API prices the small fast judge at $0.10 per million input tokens. Issue 11, coverage window 5–7 October 2026.
No China coverage this issue: the scheduled October 7 China desk run produced no branch or draft — the second pipeline disruption this week. Per EDITORIAL.md §10.6, the India desk opened this issue at 10:00 and the US desk integrated at 11:00; China’s open items carry unchanged in its October 4 ledger and were not actioned here.
Week of October 5–7, 2026 · Issue 11
The Week in One Sentence
India’s open voice line shipped the governance the commercial side still hasn’t written — cloning consent as a condition of use on an MIT model card — Google put modular multimodal retrieval on-device with a sub-1B embedder, OpenAI priced the small fast judge at a dime per million input tokens, and Claude Sonnet 4.5 got a dated retirement — a window where the publishable engineering came from two regions, a third desk’s run never happened, and the fourth’s labs shipped chores.
The Big Story: IndicF5’s update points forward — and carries consent terms
What happened. (India.) On October 6, AI4Bharat’s IndicF5 — a near-human TTS model covering 11 Indian languages, built on F5-TTS — was updated on Hugging Face (the org’s listing dated the change; the linked GitHub repo has no commits since October 1, so this was an HF-side change). The card now carries a flag-marked note, verbatim: “🚩NOTE 🚩Indic-Speak is now available. It supports 22 languages and better TTS; we recommend Indic-Speak instead.” The successor’s own page returns 401 — gated, not yet publicly readable — so everything about Indic-Speak beyond its existence, its 22-language count, and the recommendation is unverified from here.
Under the hood. IndicF5 itself, card-read in full: MIT license (field read verbatim), 0.4B parameters in F32 safetensors, gated behind login and condition-acceptance, trained on 1,417 hours across four corpora (Rasa, IndicTTS from IIT Madras, LIMMITS, IndicVoices-R). Inference is reference-voice synthesis — text plus a reference prompt audio plus that audio’s transcript — at 24 kHz, loaded with trust_remote_code=True from a GitHub pip package. The adoption numbers say this line matters beyond its quiet release cycle: 31,404 downloads last month, 281 likes, 18 community finetunes plus an adapter and a quantization, and 62 Spaces building on it. The public demo Space (INF5) runs on ZeroGPU.
What the evidence says. The consent terms are on the card, verbatim: “By using this model, you agree to only clone voices for which you have explicit permission.” Set that against the two-week arc this publication has tracked on the commercial side: Sarvam’s Voice Cloning API shipped September 26 with no consent-verification terms visible in its changelog; the vendor dropped its own cloning skill during live verification on September 28 (Issue 9); the skill has not returned through October 6, while the API itself gained streaming endpoints (see Elsewhere). One ecosystem ships governance as a condition of use; the other ships capability and hasn’t published the governance. We label both as observations, not verdicts — but the asymmetry is now quotable from primary artifacts on both sides.
Why this matters. Voice cloning is the governance question of the voice stack, and India — many languages, many voices — is where it gets decided at scale. The open line’s answer is a usage-condition gate on the model card; whether that binds anyone is untested, but it is at least written down where the weights are downloaded. The 11→22 language succession matters too: if Indic-Speak delivers all scheduled languages at IndicF5’s quality bar, the open Indic voice stack reaches parity with the hosted platforms’ headline language counts — at which point the differentiator is exactly the stuff around the model: licensing, gating, and terms.
What I would test. The reference-voice quality floor across languages: one fixed sentence set per language, three reference speakers each, measured on speaker-similarity and naturalness against the platform TTS baselines — the interesting cell is the lowest-resource languages (Assamese, Odia, Punjabi), where 1,417 training hours is thinnest. Second: whether the card’s consent terms survive into the 18 community finetunes — finetunes inherit weights, not necessarily terms, and that gap is measurable by reading finetune cards.
Elsewhere in AI Engineering
Sarvam’s voice stack, two weeks on (India) — Worth Reading
What happened. (Labelled update spanning the missed October 1–4 cycle — that issue shipped China-only in a catch-up run, so the October 1 entry below was never covered; its date is stated.) October 1, per the docs changelog: Voice Cloning gained streaming — POST /voices/clone/stream for a single binary HTTP stream, GET /voices/clone/ws for “a persistent WebSocket that accepts incremental text,” with guides and API-reference docs published. October 6 (this window): Bulbul V4 arrived for Voice Agents alongside V3 — the voice picker became a filterable catalog for browsing and previewing voices per agent, with descriptive display names (“Shubh - Hinglish Ad Voice,” “Ritu - Hindi Support Agent”) and a full V3-to-V4 name mapping. The same day, the skills repo documented the migration guidance: new integrations use the v4-flash model with a voice/persona speaker, while v3 short names still select Bulbul V3.
Why this matters. The cloning API’s endpoint surface matured — streaming is the production shape for telephony and live agents — while its official example stays dropped (checked through October 6: no cloning-skill commits since the September 28 removal) and its consent terms remain undocumented. The persona-naming work is the more quietly useful engineering: descriptive voice identities (“Hinglish Ad Voice”) turn the voice picker from an ID-lookup into a catalog, and the V3→V4 mapping plus the short-name compatibility rule is exactly the migration surface an integration team needs — the same lifecycle discipline, note, that the sarvam-m deprecation signal pointed at (Issue 9’s carry, still awaiting an official notice).
Go deeper: changelog · skills PR #24
EmbeddingGemma 2 put multimodal retrieval on-device (US) — Worth Testing
What happened. October 6, Google announced EmbeddingGemma 2 — an open, natively multimodal embedding model mapping text, code, images, video, and audio into one shared 768-dimension space. Dating per house rule: the weights have been public on HF since September 14; the announcement is this window’s event. The card, read in full: 740M parameters total, modular — 270M text-only, +170M vision, +300M audio (load just the encoders you need); Matryoshka truncation to 512/256/128 dimensions; an 8,192-token shared window (4× v1) with per-modality token costs (image 280, video frame 140, audio 25/sec); interleaved inputs via placeholder tokens; bfloat16/float32 only — the card warns float16 silently degrades.
Technical signal. Card-stated benchmarks: MTEB Code 78.68 vs 68.76 for v1 (+9.92 — the headline gain), MTEB multilingual 61.36 vs 61.15, MMEB v2 rows across image/VisDoc/video; truncation is near-lossless to 256d and drops at 128d. The blog claims ~191MB active RAM text-only (~567MB full), quantized, on a Pixel 11 Pro. Weights on HF and Kaggle, LiteRT-optimized variants, and deployment paths through MediaPipe, WebGPU, Ollama, llama.cpp, and vLLM. License caveat, flagged: the blog, card, and tag all say Apache-2.0, but the LICENSE artifact returned 404 at the repo root and the card’s license link points at Google’s Gemma license page — treat the license as declared Apache-2.0, worth verifying before commercial reliance.
Why it matters. A sub-1B multimodal embedder that runs on a phone is the missing boring layer for on-device RAG — and it lands in the same quarter the open document stack (parsing, retrieval — Issues 6–8) assembled server-side. For engineers, the modular encoders plus Matryoshka truncation make the memory/quality trade explicit at load time, which is exactly the shape edge deployment wants. First thing to test: the 128d truncation cliff on your own retrieval set before trusting the small end.
Go deeper: model card · announcement
OpenAI’s Decisions API prices the small fast judge (US) — Worth Testing
What happened. October 6, OpenAI’s Decisions API entered public beta (docs read from the primary; surfacing via HN): POST /v1/decisions “evaluates text, images, or both and returns typed answers about 10x faster than the Responses API.” Three question kinds: predicate (returns a 0–1 probability — “visible damage,” “passage relevance”), choice (routing with probabilities and confidence), and score (a probability-weighted average over your ordered severity levels). Inputs are text plus inline-base64 images; multiple independent questions ride one request. It runs only on gpt-6-luna at $0.10 per million input tokens — input-only billing, no cache or output charges (regional and long-context multipliers apply); ZDR and HIPAA eligible; GA promised “in the coming weeks”; a companion guide wires it into the Live API for voice agents.
Why it matters. Classification, routing, and triage are the volume workloads under every agent stack, and this prices them like infrastructure: a tenth of a dollar per million input tokens with no output bill, on the cheapest tier of the fleet. That is the “small fast judge” pattern productized — and its obvious use is as the cheap layer under the expensive models this issue’s ledger keeps pricing (the Decisions-route-then-reason pipeline). Watch what the ecosystem does with predicate probabilities as thresholds: the guide itself frames tuning around false-positive versus false-negative costs, which is queue-design, not prompt-design. Beta caveats stand: one model, base64-only images, dependent decisions need separate calls.
Go deeper: Decisions API guide
Anthropic’s fortnight: Sonnet 4.5 gets a retirement date, thinking gets introspectable (US) — Worth Reading
What happened. Three changelog moves, dates labelled. (September 30 — announced after this desk’s Issue 9 runs, caught up here.) Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires from the API on November 30, 2026, with Sonnet 5.5 as the replacement and a dedicated migration guide — the same entry graduated the Admin API out of beta across eight SDKs. October 1: model listings gained a line field, so fleets can group Opus/Sonnet generations without parsing ids. October 5: the Models API now reports capabilities.thinking.types.disabled per model — whether each model accepts thinking: {"type": "disabled"} is machine-readable.
Why it matters. The October 5 change quietly closes the loop on a problem this desk documented in Issue 9: thinking: "disabled" means something different on Opus 5.5 (a 400), Sonnet 5.5 (use between_tools), and earlier models — and now your code can ask the API which world it is in instead of hardcoding the matrix. Between that, the line field, and a dated retirement eight weeks out, the platform is assembling exactly the introspection and lifecycle surface a multi-model fleet needs. If you still serve Sonnet 4.5 anywhere, the migration clock reads under eight weeks.
Go deeper: changelog
No Japan items this issue: the window’s Japan-lab output was upstream-syncing serving forks (llm-jp now maintains llama.cpp, ollama, and sglang forks — PR-shaped work with no authored commits or upstream PRs verifiable yet), CI/EOL chores, and branch-metadata pushes — none cleared the editorial bar. This is the first live Japan window since Issue 9’s two items, and it produced nothing worth a reader’s time; the desk’s continuity checks re-verified unchanged (PLaMo’s license stands as read, LLM-jp-4.1 still has no independent evaluation anywhere — including NII’s own leaderboard). Japan coverage returns when the labs ship artifacts.
Research Radar
No India paper met the read-beyond-abstract bar this window (endpoint-only discovery; the search backend is down until October 13). IndicF5 ships with a 2025-era citation rather than a new paper; if Indic-Speak’s ungating brings a report, it enters here.
Repo Radar
ai4bharat/IndicF5 (India)
- What’s actually in it: 0.4B F32 safetensors (gated: login + accept), a card with license, consent terms, and training-data accounting, and a GitHub pip package for inference. Inspection depth: card read in full; files not downloaded; GitHub repo listing only.
- What you can run: reference-voice TTS for 11 languages at 24 kHz — text + reference audio + transcript in, speech out;
trust_remote_code=Trueis required, so read the custom code before you run it. - What’s missing: the successor’s cards and any new paper; per-language quality tables; and the consent terms exist only as card prose — nothing technical enforces them.
- Why open it: the reference point for open Indic voice synthesis — and the terms-bearing artifact the commercial cloning thread is currently missing.
My Read on the Week
(Synthesis and interpretation — not reported fact. Written after a cross-item pass spanning all desks’ items — two desks published this issue; China’s run never happened and Japan’s window was empty, both noted above.)
1. Governance is arriving as a card-level artifact — unevenly, and the verification burden is moving to the reader. IndicF5 ships cloning consent as a condition of use, written where the weights are downloaded; Sarvam’s cloning API still has no published consent terms two weeks in; and EmbeddingGemma 2 declares Apache-2.0 in three places while the LICENSE artifact 404s — declared, not verified, exactly the gap this publication flagged on PLaMo two issues ago (and closed only by reading the vendor’s shipped bundle). The pattern across all of it: the terms layer is thickening faster than the artifacts that prove it, so “read the license artifact” now means “notice when there isn’t one.”
2. The cheap layer under the agents got productized on both sides of the Pacific. OpenAI’s Decisions API prices classification, routing, and triage — the volume workloads under every agent stack — at $0.10 per million input tokens with no output bill; EmbeddingGemma 2 makes retrieval a sub-1B modular model that runs on a phone; Sarvam’s Bulbul V4 personas and v4-flash routing package the same boring layer for the voice side. The frontier launches take the headlines; the judge-and-embed layer is where the request volume and the invoices actually live, and this window it got explicit price tags and deployment shapes.
3. Fleet management became a platform feature. Sonnet 4.5’s dated retirement (eight weeks out), the line field for grouping model generations, and machine-readable capabilities.thinking.types.disabled together close a loop this publication documented in Issue 9 — when “thinking off” meant three different things across one vendor’s lineup, we asked for exactly this: let code ask the API which world it is in. As breaking changes become quarterly across vendors, the introspection surface stops being convenience and becomes the difference between a fleet you operate and a fleet you rewrite.
What I’d Watch Next
- Indic-Speak’s ungating — when the 401 lifts: license read (HIGH), the 22-language list, quality claims against IndicF5’s floor.
- The migration clock — Claude Sonnet 4.5 retires from the API November 30; anyone still on retired or pre-retiring models should re-pin, re-test, and re-price in one pass.
- The October 13 search reset — the cross-desk backlog clears in one cycle: China’s deferred rechecks, the PLaMo license browser-render confirmation, Sakana’s SAIL post-dating, and the first llm-jp-authored serving PRs (ollama/sglang/llama.cpp forks are syncing — the pfnet→vLLM precedent says authored PRs follow).
One Thing I’d Experiment With This Week
Price the consent gap. Run one 1,000-sentence Indic voice workload (say, Hindi + Tamil + Odia, 3 reference speakers) two ways: self-hosted IndicF5 on a single rented GPU (measure GPU-hours and speaker-similarity per language), and the hosted Bulbul v4-flash API (price the API leg from the current rate card the day you run it — we deliberately cite no prices here). The output is a three-column table: cost, quality, and governance posture (card terms vs undocumented). The third column is the one nobody’s spreadsheet has yet — and this fortnight’s artifacts make it fillable for the first time in this ecosystem.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-10-07/.
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {The Open {TTS} Line Shipped Consent Terms the Commercial One
Hasn’t},
date = {2026-10-07},
url = {https://news.llm101.in/issues/2026-10-07/},
langid = {en}
}