The listening layer: ASR, Google Chirp and Jigsaw

In January 2026, GDELT announced it had finished machine translating the Internet Archive's entire television news archive. More than 2.4 million non-English broadcasts, a quarter century of material, in collaboration with the Archive's Television News Archive. The cost they published was $54,000.

That number is the whole story of this post. Fifteen years ago, listening to the world's broadcast media was something only a government could afford, and several of them did it at enormous expense. Now it costs about the price of a mid-range car, and the reason is a layer of technology almost nobody talks about by name.

I spent a week building on GDELT for Apprised, and wrote about the dataset itself. This is the layer underneath it: the speech recognition, the models, and the handful of programs that keep the whole arrangement standing. Some of them are in better health than others.

The technology that changed shape

Automatic speech recognition turns spoken audio into text. For most of its history it did that through a pipeline of hand-built parts, and the standard description of that pipeline comes from a 2012 paper by Geoffrey Hinton and colleagues at four research groups. They wrote that "most current speech recognition systems use hidden Markov models (HMMs) to deal with the temporal variability of speech and Gaussian mixture models to determine how well each state of each HMM fits a frame or a short window of frames of coefficients that represents the acoustic input."

Their contribution was to replace one component of that stack, the Gaussian scoring step, with a deep neural network. It worked well enough that they could report neural networks outperforming the old approach "on a variety of speech recognition benchmarks, sometimes by a large margin."

Two years later, Baidu's Deep Speech paper threw the rest of the pipeline away. Its authors were blunt about what was being discarded: "We do not need a phoneme dictionary, nor even the concept of a 'phoneme.'" No hand-designed noise model, no separate reverberation handling, no linguistics. Audio in, text out, learned end to end.

Everything since has been a matter of scale.

How we measure it, and what the benchmark quietly encodes

The field's metric is word error rate. The reference implementation is sclite, part of the Speech Recognition Scoring Toolkit that NIST has maintained since 2009 and updated as recently as August 2026.

Here is a small thing worth knowing. Nearly every paper attributes the word error rate formula to NIST, and sclite's own documentation does not actually print it. What the docs define are separate ratios, percent correct, percent substituted, percent inserted, percent deleted, each computed over the number of reference words after a dynamic programming alignment. The single combined figure everyone quotes is an aggregation the community settled on, not a definition the toolkit hands you.

The corpus those scores are usually quoted against is LibriSpeech: a thousand hours of read English drawn from LibriVox public domain audiobooks, built at Johns Hopkins and released under CC BY 4.0. Its test sets are split into "clean" and "other," and thousands of papers report both numbers.

What "clean" means is more interesting than it sounds. The authors describe selecting the first sets to be "on average, of higher recording quality and with accents closer to US English," and they operationalised both properties through a single number: the word error rate of an acoustic model trained on the Wall Street Journal corpus. Speakers the old model found easy went in one pile, speakers it found hard went in the other. For the harder sets they say they "deliberately chose more challenging data."

So the field's canonical easy versus hard axis bundles recording quality and accent proximity together, by construction, behind a word that sounds purely acoustic. That is not a scandal, the paper says exactly what it did. It is just worth knowing when you read a benchmark table.

Chirp, and the model underneath it

Google's speech recognition brand on Cloud is Chirp. The research it came from is the Universal Speech Model, published in March 2023 as "Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages".

The training scale is the part worth repeating. USM was pre-trained on YT-NTL-U, described in the paper as 12 million hours of YouTube audio covering over 300 languages, alongside a text corpus of 28 billion sentences across more than 1,140 languages. The models are 2 billion parameter Conformers, a figure the abstract omits and the body states. Against Whisper, the paper claims comparable or better quality "while using less than 1/7-th of the amount of labeled training data."

The language counts are where it gets muddy, and the muddiness is instructive.

Google's 1,000 Languages Initiative, announced 2 November 2022, described a model "trained on over 400 languages." The paper four months later leads with 100 and describes pre-training across 300. The research blog calls the work "a critical first step towards supporting 1,000 languages." Inside the paper itself, a summary bullet describes a 90,000 hour supervised set as covering 73 languages, while Section 3.1 describes what appears to be the same 90,000 hours as spanning 75. These are not lies, they are different slices of different datasets reported in different places, but you cannot reconcile them from the outside, and nobody publishes a single authoritative figure.

The current documentation is stranger still. The Chirp 3 page, last updated in July 2026, lists 111 languages in its transcription table. The Speech-to-Text product page carries both "Supports over 125 languages" and "Support for 85+ languages and variants" in different sections of the same page.

And the research lineage has been quietly scrubbed. I counted occurrences of the string "USM" across four current Speech-to-Text pages: the Chirp 2 page, the Chirp 3 page, the supported languages matrix, and the model comparison table. Three of them contain it zero times. The old chirp-model documentation URL, which used to be titled "Chirp: Universal speech model" and linked the USM paper directly, now redirects and then 404s. I checked that from this machine on 8 August 2026.

One trap for anyone searching: "Chirp 3: HD voices" is a Text-to-Speech product. It is a different thing from Chirp 3 speech recognition, and Google ships both under the same name.

Jigsaw, and the thing that is ending

If you have read GDELT's front page you have seen Jigsaw's name without necessarily registering it. The site says, verbatim, "Supported by Google Jigsaw, the GDELT Project monitors the world's broadcast, print, and web news."

Jigsaw describes itself as "an incubator within Google that builds technologies to give people greater agency in the world around them." It began as Google Ideas. Eric Schmidt announced the change in February 2016: "Today we're announcing the expansion of Google Ideas, Google's think tank, as a technology incubator called Jigsaw." The original mission language was geopolitical, "from countering violent extremism to thwarting online censorship to mitigating the threats associated with digital attacks."

The name Google Ideas now appears zero times anywhere on Jigsaw's current site. I crawled all eight pages in its sitemap to check.

Its best known product is Perspective API, which scores text for the perceived impact a comment may have on a conversation. Six production attributes: TOXICITY, SEVERE_TOXICITY, IDENTITY_ATTACK, INSULT, PROFANITY and THREAT. It has always been free, with a default quota of one query per second. Jigsaw's own page says that "since launching in 2017, the API has been used by over 1,000 partners in 18 languages almost two billion times a day," and credits it with letting the New York Times "triple the number of articles that are available for commenting."

Perspective API is being switched off. Its home page now carries a notice that "Perspective API is sunsetting and service is officially ending after 2026." The service remains active until 31 December 2026. Usage requests were handled only until February 2026. The stated reason is that "AI capabilities have evolved, and there is now less demand for a standalone tool specific to this area." Asked whether migration support will be offered, the FAQ answers: "No. We will not be offering direct migration support, but there are various tools available that were inspired by Perspective's capabilities."

Something used two billion times a day is being retired with no migration path, because general purpose models absorbed the use case.

The rest of the portfolio is mixed rather than uniformly dismantled, and it is worth being precise about that. Of seven named products, Perspective is sunsetting, Tune's repository states plainly that it "is no longer available," and Assembler's subdomain no longer resolves at all, a DNS lookup returns NXDOMAIN. Altitude has moved to Tech Against Terrorism. Project Shield, launched in 2013 as free denial of service protection for news and elections sites, is still running. Outline is still running, though it is now owned by the Outline Foundation, described on its own site as "an independent non-profit organization," rather than by Google.

Jigsaw also co-authored the main peer reviewed work on prebunking, "Psychological inoculation improves resilience against misinformation on social media," in Science Advances, August 2022.

What the listening actually feeds

The Internet Archive's TV News Archive has been running since September 2012. Its live counter read 4,325,000 shows when I checked it on 8 August 2026. Its public search box is scoped, in its own placeholder text, to "Search captions (2009 to yesterday)." Shareable quotes are capped at 60 seconds. In 2014 the Archive wrote that it "does not sell or license this content," and offered researchers a full show on DVD-ROM for a $25 processing fee; today the site offers a USB flash drive with no price stated.

GDELT's Visual Explorer sits on top of that collection, and GDELT reproduced a line from the Archive that stuck with me: "It's also a hidden gem, with only 20,000 or so unique visitors a day." A complete searchable record of American and international television, and the daily audience of a mid-sized newsletter.

Compare that with what the same capability used to cost and who could have it.

BBC Monitoring was established in 1939, in Parliament's own words, "to listen in on World War II propaganda broadcasts." Until 2013 the bulk of its funding came from the Foreign and Commonwealth Office, the Ministry of Defence and the Cabinet Office, with responsibility transferring to the BBC on 1 April 2013. Its output today sits behind monitoring.bbc.co.uk, a subscriber portal that serves a JavaScript shell and nothing else to an anonymous visitor.

The American equivalent went the other way and then closed. The Foreign Broadcast Information Service ran a public product, the World News Connection, for decades. The CIA terminated public access at the end of 2013. The Open Source Center became the Open Source Enterprise on 1 October 2015, and a CIA spokesperson said at the time that it had "no plans to expand the scope of its services to include the regular release of unclassified, non-copyrighted materials to the public." The historical FBIS reports from 1941 to 1996 are available commercially through Readex, at a price you have to ask for.

So the public listening layer is not an extension of the government one. It replaced it, from a completely different direction, and it is run by a digital library, a university researcher, a search company's incubator and a handful of open model releases.

The programs almost nobody names

Beyond Chirp and Whisper, the field runs on work most people building with speech have never heard of.

Meta's Massively Multilingual Speech is the widest coverage anyone has published: pre-trained models supporting 1,107 languages for recognition and language identification across 4,017. It is also licensed CC-BY-NC 4.0, which means the widest multilingual speech models in existence are not available for commercial use. The repository that hosts them, fairseq, was archived in September 2025.

NVIDIA's NeMo family is the quiet workhorse of open ASR. Canary-1b-v2 is a 978 million parameter encoder-decoder covering 25 European languages; Parakeet-TDT-0.6b-v3 is 600 million parameters with automatic language detection across the same set. Both are CC-BY-4.0, which is to say genuinely usable commercially. NVIDIA's own docs are careful to describe Parakeet as "a family of ASR models," not one model, and the main repository has moved from NVIDIA/NeMo to NVIDIA-NeMo/Speech.

Coqui STT is the cautionary tale. It was the community successor to Mozilla's DeepSpeech, and it is dead. Its README now directs users to Whisper, and coqui.ai returns a GitHub Pages 404. The repository has never been marked archived, which means a newcomer searching for open speech recognition still finds a live-looking project with recent-looking stars.

Mozilla Common Voice is the largest public donated speech corpus, and it moved house in a way most practitioners missed. As of October 2025 the datasets are no longer distributed through Hugging Face; the old repositories are empty stubs, and distribution runs through Mozilla Data Collective, which describes itself as "a mission-locked British company, incubated by Mozilla Foundation." The Scripted Speech release v26.0, dated June 2026, holds 42,388 total hours across 294 languages, of which 28,893 are validated.

Common Voice also contains the single most surprising number I found. Pulling its live language statistics API and counting, the corpus covers 441 locales, but only 282 have any validated audio at all and only 39 have more than a hundred validated hours. And the largest validated corpus in the entire project is not English. It is Pashto, at 3,520 hours, ahead of English at 2,783.

That is what a donation-driven corpus looks like when a determined community shows up. It is also a reminder that "supports 441 languages" and "usefully supports 39 languages" are both true statements about the same dataset.

Two more worth naming. CLARIN ERIC is a European research infrastructure with actual legal personality under EC Regulation 723/2009, with its statutory seat in Utrecht, holding language resources for the humanities. And Te Hiku Media, a Māori broadcaster in New Zealand, built its own speech recognition for te reo Māori and released it under the Kaitiakitanga License, which forbids commercial use without permission. When a community records its own language, it does not necessarily want the result to become training data for everyone.

The limits, stated by the people who built it

None of this works as well as the marketing implies, and the honest numbers come from the vendors themselves.

OpenAI's Whisper model card states that the models "show strong ASR results in ~10 languages," and that they "exhibit disparate performance on different accents and dialects of particular languages." Whisper was trained on 680,000 hours and ships in six sizes under an MIT licence, and its own documentation will tell you it is uneven.

The independent measurement is worse. Koenecke and colleagues, in PNAS, tested five commercial speech recognition systems and found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers. Same audio conditions, same task, roughly double the error rate.

And performance collapses outside clean recordings. In the CHiME-7 challenge on real meeting audio, Whisper large scored 53.2% macro-averaged diarization-attributed word error rate on the evaluation set in the main track, and 32.5% in the acoustic robustness sub-track. That metric charges the system for getting speaker attribution wrong as well as words, which is the right thing to measure and a much harder test than reading audiobooks.

A tool that transcribes a podcast flawlessly can fail badly on a noisy multi-speaker broadcast in an accent it saw little of. Every media monitoring claim built on ASR inherits that.

Why this is worth your attention

The thing I keep returning to, from the GDELT post and now from this one, is how contingent the arrangement is.

The listening layer is real. It genuinely lets one person with a laptop do something that used to require a government agency, and $54,000 for a quarter century of translated world television is not a rounding error on that claim, it is the proof.

But it is not stable, and the pieces do not announce their departure. Perspective API is going after 2026 with no migration path. Coqui is dead and does not look dead. Common Voice changed distributor and most people building on it never noticed. The widest multilingual speech models cannot be used commercially, and their repository is archived. BBC Monitoring went behind a subscription; the CIA's public feed closed in 2013 and never reopened.

What survives is what someone chooses to keep running. GDELT survives because a researcher and a Google incubator decided it should. The TV News Archive survives because a library decided it should, for twenty thousand visitors a day.

If you are building on any of this, the practical lesson is the one I keep writing: know which pieces you actually depend on, and know what you would do the week one of them is switched off. That is not pessimism. It is just the maintenance cost of standing on infrastructure you did not pay for.

Fact-check notes and sources

HMM and Gaussian mixture description, and neural networks outperforming them: Hinton, Deng, Yu, Dahl, Mohamed, Jaitly, Senior, Vanhoucke, Nguyen, Sainath and Kingsbury, "Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups," IEEE Signal Processing Magazine, 2012, abstract via Google Research
"We do not need a phoneme dictionary, nor even the concept of a 'phoneme'": Hannun et al., "Deep Speech: Scaling up end-to-end speech recognition", December 2014
SCTK and sclite, and the absence of a combined WER formula in its documentation: NIST Multimodal Information Group tools (created December 2009, updated August 2026) and the sclite documentation. The absence is my own check of that document, not a NIST statement.
LibriSpeech size, source, licence, and the clean/other selection procedure including "higher recording quality and with accents closer to US English" and "deliberately chose more challenging data": Panayotov, Chen, Povey and Khudanpur, "LibriSpeech: an ASR corpus based on public domain audio books", ICASSP 2015, and OpenSLR 12
USM training scale, 2 billion parameter Conformers, the 1/7th labeled data comparison, and the 73 versus 75 language figures: "Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages", March 2023
"a critical first step towards supporting 1,000 languages": Google Research blog on USM
1,000 Languages Initiative and "trained on over 400 languages": Google blog, 2 November 2022
Chirp announced at Google I/O and detailed on 19 May 2023: Google Cloud blog
Current Chirp 3 language table row count, the two conflicting marketing counts, the absence of "USM" from three of four current Chirp-related documentation pages, and the chirp-model redirect ending in a 404: my own measurements against Google Cloud documentation on 8 August 2026, not Google statements.
Jigsaw self-description, structural agency language and the "fifteen years ago" line: jigsaw.google. Google Ideas to Jigsaw announcement and the original mission language: Eric Schmidt, February 2016, via the Internet Archive snapshot
Perspective API sunset, the December 2026 end date, the stated reason and the absence of migration support: perspectiveapi.com. Production attributes, free access and the one query per second default: Perspective developer documentation
"over 1,000 partners in 18 languages almost two billion times a day" and the New York Times commenting figure: Jigsaw, Supporting Online Conversations
Prebunking research: "Psychological inoculation improves resilience against misinformation on social media," Science Advances, August 2022, volume 8 number 34, DOI 10.1126/sciadv.abo6254
Tune "is no longer available": its GitHub README. Outline Foundation ownership: getoutline.org. Project Shield launch year and remit: Jigsaw, Protecting Essential Information. The Assembler subdomain returning NXDOMAIN is my own DNS check on 8 August 2026.
TV News Archive launch date, the 60-second quote cap and the 2014 DVD-ROM terms: Internet Archive blog, 1 April 2014. The 4,325,000 show count is the live value at av.archive.org/etc/info.json on 8 August 2026.
$54,000 translation of more than 2.4 million non-English broadcasts, in collaboration with the Internet Archive's Television News Archive: GDELT, 22 January 2026. The "hidden gem, with only 20,000 or so unique visitors a day" line is reproduced by GDELT and attributed there to the Internet Archive.
BBC Monitoring founded 1939, its pre-2013 government funders and the 1 April 2013 transfer: House of Commons Foreign Affairs Committee, "The future operations of BBC Monitoring," HC 732 (2016-17)
CIA termination of public access at the end of 2013, and the Open Source Center becoming the Open Source Enterprise on 1 October 2015 including the spokesperson quote: Steven Aftergood, Federation of American Scientists, January 2014 and October 2015. These are secondary sources; the CIA has published no primary page on either change.
MMS language counts and licence: "Scaling Speech Technology to 1,000+ Languages" and the fairseq MMS README. The fairseq archive date is from the GitHub API on 8 August 2026.
NVIDIA Canary-1b-v2 and Parakeet-TDT-0.6b-v3 parameter counts, language coverage and CC-BY-4.0 terms: their Hugging Face model cards. "Parakeet is the name of a family of ASR models": NVIDIA NeMo documentation
Coqui STT status and the Whisper redirect: its GitHub README. The coqui.ai 404 and the repository's un-archived state are my own checks on 8 August 2026.
Common Voice distribution move, Mozilla Data Collective's self-description, and the v26.0 Scripted Speech hours: Common Voice terms, Mozilla Data Collective and the cv-dataset release table. The 441 locales, 282 with validated audio, 39 above a hundred hours, and Pashto at 3,520 hours ahead of English at 2,783, are my own counts against the live Common Voice statistics API on 8 August 2026.
CLARIN ERIC's legal basis and statutory seat: its statutes. Te Hiku Media's Kaitiakitanga License terms: the licence text
Whisper training hours, model sizes, licence, "~10 languages" and "disparate performance on different accents and dialects": the Whisper repository and its model card
0.35 versus 0.19 word error rate by speaker race across five commercial systems: Koenecke et al., "Racial disparities in automated speech recognition," PNAS, 2020, read via PubMed Central because pnas.org refused requests from my address
Whisper large CHiME-7 figures: Table 3 of "The CHiME-7 DASR Challenge". The 53.2 figure is the evaluation set, main track, macro average; the sub-track figure is 32.5.

Where I have measured something myself rather than quoting a source, I have said so above. Language counts in this field move constantly, and every figure carries the date I checked it.

The listening layer: ASR, Google Chirp, Jigsaw and the programs holding up global media monitoring