# SpeechData.ai

> SpeechData.ai licenses conversational speech datasets in 60 languages and dialects for training ASR, TTS and voice AI models, and collects custom speech data to spec. It is part of Spirelight (https://www.spirelight.ai).

## Buying facts

- Catalog: 60 off-the-shelf datasets, one per language or dialect, 73,000 hours from 7,250 vetted native speakers
- Speech: spontaneous two-person conversations recorded in-country, not read scripts
- Audio: dual-channel, one channel per speaker
- Transcripts: time-aligned and human-verified
- Metadata: age, gender, region and recording device for every speaker
- Consent: every speaker signed explicit commercial-use consent; the full consent chain is provided for legal review
- License: commercial use, including training and evaluating ASR, TTS and other voice AI models
- Price: $60 to $95 per hour of audio, published on each dataset page; volume and multi-language discounts on request; prices by region: https://www.speechdata.ai/speech-data-pricing
- Samples: free sample audio, matching transcripts and the metadata schema, sent on request to a company email
- Custom collection: languages, dialects, domains, devices or acoustic environments outside the catalog, recorded to spec as sentence or prompt lists, monologues or spontaneous two-person dialogue; native speakers recruited to quotas for age, gender, region and accent, identity, hardware and background noise checked before recording; feasibility and a quote within one business day
- TTS and voice cloning data: native speakers record sentence lists, prompts or monologues, and the buyer can approve each voice from a short sample; each speaker is briefed on how the recordings will be used and signs a speaker contract for that project; quoted per project
- Wake word data: wake phrases, commands and confusable phrases from the buyer's prompt list, recorded by native speakers to accent and demographic quotas, on phone, headset, far-field or in-car audio; quoted per project
- Accent and dialect data: the catalog has English from the United States, South Africa and Singapore; Spanish from Spain and Mexico; French from France and the DR Congo; Portuguese from Portugal, Brazil and Angola; Dutch from the Netherlands and Suriname; and Arabic in 5 varieties (Modern Standard, Gulf, Levantine, Egyptian and Darija), each variety its own dataset recorded in-country by native speakers, with every speaker's region in the metadata; other accents and varieties are recorded by native speakers to quotas for region, accent, age and gender after a feasibility check, quoted per project
- In-car speech data: wake words, commands, navigation and media requests, addresses, monologues or dialogue recorded by native speakers to quotas for age, gender, region and accent, with in-car audio as the brief specifies; the brief can set the speaker's seat, microphone placement, whether the car is parked, idling or driving, windows, ventilation, media and cabin noise, each covered in the feasibility assessment; quoted per project
- Scripted speech data: the buyer's sentence or prompt lists (product and brand names, people and place names, addresses, numbers, dates, amounts, codes, domain terms, commands) read by native speakers recruited to quotas for age, gender, region and accent, on the devices and in the environments specified; time-aligned transcripts in the buyer's convention, clean-read or verbatim with disfluencies tagged; the catalog is spontaneous conversation only, so scripted speech is always recorded to spec; quoted per project
- Call center data: role-played customer service calls recorded to spec by native speakers playing agent and customer, from industry scenario scripts, one channel per speaker, narrowband telephony or wideband; fictional names and account details, so no real customer data; transcripts with speaker turns and timestamps; quoted per project
- Annotation: transcription and labeling of the buyer's own audio as a standalone service (verbatim transcripts, timestamps, speaker turns, non-speech events, intent, sentiment and emotion, custom schemas) by native-speaker annotators under a written guideline, delivered as JSON, JSONL, CSV, SRT, VTT or a custom manifest; priced per audio hour, item or batch after reviewing a sample
- South African languages: South African English off the shelf (2,000 hours); isiZulu, isiXhosa and Sesotho recorded to spec by native speakers; other South African languages on request

## Services

- [Multilingual speech data collection services](https://www.speechdata.ai/custom-speech-data-collection): custom audio and speech data collection: scripted or spontaneous speech, target demographics, devices and acoustic environments, recruited, recorded, QA'd, transcribed and delivered
- [TTS and voice cloning data collection](https://www.speechdata.ai/tts-data-collection): text-to-speech and voice cloning training data recorded by native speakers as sentence lists, prompts or monologues, with voices the buyer can approve from a sample, checked and transcribed
- [Wake word data collection](https://www.speechdata.ai/wake-word-data-collection): wake word, voice command and confusable phrase recordings by native speakers to accent quotas, on the devices and in the environments the product will meet
- [Accent and dialect speech data](https://www.speechdata.ai/accent-speech-data-collection): 6 languages in 17 varieties off the shelf, each recorded in-country, and native speakers recruited to region and accent quotas for accents outside the catalog
- [In-car speech data collection](https://www.speechdata.ai/in-car-speech-data-collection): wake words, commands, navigation and media requests and dialogue recorded by native speakers to quotas, with the speaker's seat, microphone placement and cabin conditions set in the brief
- [Scripted speech data collection](https://www.speechdata.ai/scripted-speech-data-collection): sentence and prompt lists of names, addresses, numbers, codes, domain terms and commands read by many native speakers to quotas, checked and transcribed in the buyer's convention
- [Call center speech data](https://www.speechdata.ai/call-center-speech-data): role-played customer service calls in industry scenarios, recorded by native speakers on separate agent and customer channels, with transcripts and optional intent, outcome, sentiment and PII labels
- [Audio annotation and transcription services](https://www.speechdata.ai/audio-annotation-services): transcription, timestamps, speaker turns, event, intent and emotion labels on the buyer's own audio by native-speaker annotators, with a written guideline and measured QA
- [South African speech data](https://www.speechdata.ai/south-african-speech-data): South African English dataset off the shelf; isiZulu, isiXhosa and Sesotho speech data recorded to spec by native speakers
- [Speech dataset catalog](https://www.speechdata.ai/datasets): all 60 off-the-shelf conversational speech datasets by region
- [Speech data pricing](https://www.speechdata.ai/speech-data-pricing): published prices per hour of audio for all 60 datasets, by region, with worked budget examples; custom collection, call-center recordings and audio annotation are quoted per project
- [Request samples](https://www.speechdata.ai/contact): ask for sample audio, transcripts and the metadata schema
- [Book a call](https://www.speechdata.ai/book): 30-minute call with the founder about a data requirement

## Datasets

- [Arabic speech data: Modern Standard Arabic and 4 dialects](https://www.speechdata.ai/arabic-speech-data): Modern Standard Arabic, Gulf Arabic, Levantine Arabic, Egyptian Arabic and Moroccan Darija compared side by side: 6,500 hours from 650 native speakers recorded in Saudi Arabia, Lebanon, Egypt and Morocco, $65 to $90 per hour of audio
- [Speech data for 10 Indian languages](https://www.speechdata.ai/indian-languages-speech-data): Hindi, Bengali, Urdu, Punjabi, Gujarati, Marathi, Tamil, Telugu, Kannada and Malayalam compared side by side: 8,500 hours from 850 native speakers recorded in India, Bangladesh and Pakistan, $65 to $75 per hour of audio
- [African speech data from 8 countries](https://www.speechdata.ai/african-speech-data): Swahili, Hausa, Yoruba and Amharic plus South African English, Congolese French, Angolan Portuguese, Egyptian Arabic and Moroccan Darija, compared side by side: 16,000 hours from 1,600 native speakers recorded in Kenya, Nigeria, Ethiopia, South Africa, the DR Congo, Angola, Egypt and Morocco, $60 to $65 per hour of audio
- [Southeast Asian speech data in 6 languages](https://www.speechdata.ai/southeast-asian-speech-data): Indonesian, Malay, Thai, Vietnamese, Tagalog and Singapore English compared side by side: 12,000 hours from 1,200 native speakers recorded in Indonesia, Malaysia, Thailand, Vietnam, the Philippines and Singapore, $60 to $75 per hour of audio
- [European speech data in 22 languages](https://www.speechdata.ai/european-speech-data): European Spanish, French, European Portuguese, German, Dutch, Italian, Greek, Catalan, Swedish, Norwegian, Danish, Finnish, Polish, Russian, Ukrainian, Romanian, Czech, Slovak, Hungarian, Bulgarian, Croatian and Serbian compared side by side: 21,500 hours from 2,100 native speakers recorded in Spain, France, Portugal, Germany, the Netherlands, Italy, Greece, Sweden, Norway, Denmark, Finland, Poland, Russia, Ukraine, Romania, Czechia, Slovakia, Hungary, Bulgaria, Croatia and Serbia, $75 to $95 per hour of audio
- [East Asian speech data: Taiwanese Mandarin, Japanese and Korean](https://www.speechdata.ai/east-asian-speech-data): Taiwanese Mandarin, Japanese and Korean compared side by side: 2,500 hours from 250 native speakers recorded in Taiwan, Japan and South Korea, $70 to $95 per hour of audio
- [English (US) conversational speech dataset](https://www.speechdata.ai/datasets/english-usa): 1,500 hours from 150 native speakers recorded in the United States, $90 per hour of audio
- [English (South African) conversational speech dataset](https://www.speechdata.ai/datasets/english-south-africa): 2,000 hours from 200 native speakers recorded in South Africa, $60 per hour of audio
- [English (Singaporean) conversational speech dataset](https://www.speechdata.ai/datasets/english-singapore): 2,000 hours from 200 native speakers recorded in Singapore, $60 per hour of audio
- [Spanish (European) conversational speech dataset](https://www.speechdata.ai/datasets/spanish-spain): 1,000 hours from 100 native speakers recorded in Spain, $85 per hour of audio
- [Spanish (Latin American) conversational speech dataset](https://www.speechdata.ai/datasets/spanish-mexico): 2,000 hours from 200 native speakers recorded in Mexico, $65 per hour of audio
- [French (European) conversational speech dataset](https://www.speechdata.ai/datasets/french-france): 1,000 hours from 100 native speakers recorded in France, $95 per hour of audio
- [French (African) conversational speech dataset](https://www.speechdata.ai/datasets/french-dr-congo): 2,000 hours from 200 native speakers recorded in the DR Congo, $65 per hour of audio
- [Portuguese (European) conversational speech dataset](https://www.speechdata.ai/datasets/portuguese-portugal): 1,000 hours from 100 native speakers recorded in Portugal, $85 per hour of audio
- [Portuguese (Brazilian) conversational speech dataset](https://www.speechdata.ai/datasets/portuguese-brazil): 2,000 hours from 200 native speakers recorded in Brazil, $65 per hour of audio
- [Portuguese (African) conversational speech dataset](https://www.speechdata.ai/datasets/portuguese-angola): 500 hours from 50 native speakers recorded in Angola, $65 per hour of audio
- [German conversational speech dataset](https://www.speechdata.ai/datasets/german): 1,000 hours from 50 native speakers recorded in Germany, $95 per hour of audio
- [Dutch conversational speech dataset](https://www.speechdata.ai/datasets/dutch-netherlands): 1,000 hours from 100 native speakers recorded in the Netherlands, $95 per hour of audio
- [Dutch (Surinamese) conversational speech dataset](https://www.speechdata.ai/datasets/dutch-suriname): 1,000 hours from 100 native speakers recorded in Suriname, $65 per hour of audio
- [Italian conversational speech dataset](https://www.speechdata.ai/datasets/italian): 1,000 hours from 100 native speakers recorded in Italy, $75 per hour of audio
- [Polish conversational speech dataset](https://www.speechdata.ai/datasets/polish): 1,000 hours from 100 native speakers recorded in Poland, $85 per hour of audio
- [Russian conversational speech dataset](https://www.speechdata.ai/datasets/russian): 1,000 hours from 100 native speakers recorded in Russia, $85 per hour of audio
- [Ukrainian conversational speech dataset](https://www.speechdata.ai/datasets/ukrainian): 1,500 hours from 150 native speakers recorded in Ukraine, $85 per hour of audio
- [Romanian conversational speech dataset](https://www.speechdata.ai/datasets/romanian): 1,000 hours from 100 native speakers recorded in Romania, $85 per hour of audio
- [Greek conversational speech dataset](https://www.speechdata.ai/datasets/greek): 1,000 hours from 100 native speakers recorded in Greece, $85 per hour of audio
- [Czech conversational speech dataset](https://www.speechdata.ai/datasets/czech): 1,000 hours from 100 native speakers recorded in Czechia, $90 per hour of audio
- [Slovak conversational speech dataset](https://www.speechdata.ai/datasets/slovak): 1,000 hours from 100 native speakers recorded in Slovakia, $90 per hour of audio
- [Hungarian conversational speech dataset](https://www.speechdata.ai/datasets/hungarian): 1,000 hours from 100 native speakers recorded in Hungary, $90 per hour of audio
- [Bulgarian conversational speech dataset](https://www.speechdata.ai/datasets/bulgarian): 1,500 hours from 150 native speakers recorded in Bulgaria, $85 per hour of audio
- [Croatian conversational speech dataset](https://www.speechdata.ai/datasets/croatian): 1,500 hours from 150 native speakers recorded in Croatia, $75 per hour of audio
- [Serbian conversational speech dataset](https://www.speechdata.ai/datasets/serbian): 1,500 hours from 150 native speakers recorded in Serbia, $75 per hour of audio
- [Catalan conversational speech dataset](https://www.speechdata.ai/datasets/catalan): 500 hours from 50 native speakers recorded in Spain, $75 per hour of audio
- [Swedish conversational speech dataset](https://www.speechdata.ai/datasets/swedish): 500 hours from 50 native speakers recorded in Sweden, $95 per hour of audio
- [Norwegian conversational speech dataset](https://www.speechdata.ai/datasets/norwegian): 500 hours from 50 native speakers recorded in Norway, $95 per hour of audio
- [Danish conversational speech dataset](https://www.speechdata.ai/datasets/danish): 500 hours from 50 native speakers recorded in Denmark, $95 per hour of audio
- [Finnish conversational speech dataset](https://www.speechdata.ai/datasets/finnish): 500 hours from 50 native speakers recorded in Finland, $95 per hour of audio
- [Hindi conversational speech dataset](https://www.speechdata.ai/datasets/hindi): 2,000 hours from 200 native speakers recorded in India, $65 per hour of audio
- [Bengali conversational speech dataset](https://www.speechdata.ai/datasets/bengali): 2,000 hours from 200 native speakers recorded in Bangladesh, $65 per hour of audio
- [Urdu conversational speech dataset](https://www.speechdata.ai/datasets/urdu): 1,000 hours from 100 native speakers recorded in Pakistan, $65 per hour of audio
- [Punjabi conversational speech dataset](https://www.speechdata.ai/datasets/punjabi): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Gujarati conversational speech dataset](https://www.speechdata.ai/datasets/gujarati): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Marathi conversational speech dataset](https://www.speechdata.ai/datasets/marathi): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Tamil conversational speech dataset](https://www.speechdata.ai/datasets/tamil): 500 hours from 50 native speakers recorded in India, $75 per hour of audio
- [Telugu conversational speech dataset](https://www.speechdata.ai/datasets/telugu): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Kannada conversational speech dataset](https://www.speechdata.ai/datasets/kannada): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Malayalam conversational speech dataset](https://www.speechdata.ai/datasets/malayalam): 500 hours from 50 native speakers recorded in India, $65 per hour of audio
- [Indonesian conversational speech dataset](https://www.speechdata.ai/datasets/indonesian): 2,000 hours from 200 native speakers recorded in Indonesia, $65 per hour of audio
- [Malay conversational speech dataset](https://www.speechdata.ai/datasets/malay): 2,000 hours from 200 native speakers recorded in Malaysia, $65 per hour of audio
- [Thai conversational speech dataset](https://www.speechdata.ai/datasets/thai): 2,000 hours from 200 native speakers recorded in Thailand, $75 per hour of audio
- [Vietnamese conversational speech dataset](https://www.speechdata.ai/datasets/vietnamese): 2,000 hours from 200 native speakers recorded in Vietnam, $65 per hour of audio
- [Tagalog conversational speech dataset](https://www.speechdata.ai/datasets/tagalog): 2,000 hours from 200 native speakers recorded in the Philippines, $65 per hour of audio
- [Mandarin Chinese (Taiwanese) conversational speech dataset](https://www.speechdata.ai/datasets/mandarin-taiwan): 1,500 hours from 150 native speakers recorded in Taiwan, $70 per hour of audio
- [Japanese conversational speech dataset](https://www.speechdata.ai/datasets/japanese): 500 hours from 50 native speakers recorded in Japan, $95 per hour of audio
- [Korean conversational speech dataset](https://www.speechdata.ai/datasets/korean): 500 hours from 50 native speakers recorded in South Korea, $95 per hour of audio
- [Arabic (Modern Standard) conversational speech dataset](https://www.speechdata.ai/datasets/arabic-msa): 2,000 hours from 200 native speakers recorded in Saudi Arabia, $90 per hour of audio
- [Arabic (Gulf) conversational speech dataset](https://www.speechdata.ai/datasets/arabic-gulf): 500 hours from 50 native speakers recorded in Saudi Arabia, $90 per hour of audio
- [Arabic (Levantine) conversational speech dataset](https://www.speechdata.ai/datasets/arabic-levantine): 500 hours from 50 native speakers recorded in Lebanon, $65 per hour of audio
- [Arabic (Egyptian) conversational speech dataset](https://www.speechdata.ai/datasets/arabic-egyptian): 2,000 hours from 200 native speakers recorded in Egypt, $65 per hour of audio
- [Arabic (Darija) conversational speech dataset](https://www.speechdata.ai/datasets/arabic-darija): 1,500 hours from 150 native speakers recorded in Morocco, $65 per hour of audio
- [Hebrew conversational speech dataset](https://www.speechdata.ai/datasets/hebrew): 500 hours from 50 native speakers recorded in Israel, $95 per hour of audio
- [Persian (Farsi) conversational speech dataset](https://www.speechdata.ai/datasets/persian-farsi): 500 hours from 50 native speakers recorded in Iran, $65 per hour of audio
- [Turkish conversational speech dataset](https://www.speechdata.ai/datasets/turkish): 2,000 hours from 200 native speakers recorded in Turkey, $75 per hour of audio
- [Swahili conversational speech dataset](https://www.speechdata.ai/datasets/swahili): 2,000 hours from 200 native speakers recorded in Kenya, $65 per hour of audio
- [Hausa conversational speech dataset](https://www.speechdata.ai/datasets/hausa): 2,000 hours from 200 native speakers recorded in Nigeria, $65 per hour of audio
- [Yoruba conversational speech dataset](https://www.speechdata.ai/datasets/yoruba): 2,000 hours from 200 native speakers recorded in Nigeria, $65 per hour of audio
- [Amharic conversational speech dataset](https://www.speechdata.ai/datasets/amharic): 2,000 hours from 200 native speakers recorded in Ethiopia, $65 per hour of audio

## Buyer guides

- [Wake Word Training Data: How Much You Need and What to Record](https://www.speechdata.ai/guides/wake-word-training-data): Plan wake word training data: positives, confusable negatives and a test set, what open datasets show about scale, and a collection spec to paste.
- [Voice Command Data: How to Design the Utterance List](https://www.speechdata.ai/guides/voice-command-data): Plan voice command data collection: intents, phrasings, slot values, out-of-scope requests, speakers, devices, labels and a test set, with a spec to paste.
- [Multilingual ASR Training Data: How to Plan Each Language](https://www.speechdata.ai/guides/multilingual-asr-data): How to plan multilingual speech data for ASR: one model or several, hours per language, balancing, shared conventions, per-language tests and a budget.
- [ASR Training Data: What It Contains and How to Check It](https://www.speechdata.ai/guides/asr-training-data): What ASR training data contains (audio, transcripts, timestamps, speaker metadata, rights) and how to check a vendor's sample before you buy.
- [Voice Agent Evaluation: Metrics, Test Sets and Caller Audio](https://www.speechdata.ai/guides/voice-agent-evaluation): How to evaluate a voice agent layer by layer, from speech recognition and latency to turn-taking, task success and TTS, and which caller audio to test on.
- [Best Sample Rate for Speech Data: 8, 16, 24 or 48 kHz?](https://www.speechdata.ai/guides/speech-data-sample-rate): Which sample rate, bit depth, channel layout and format to specify for ASR, TTS and voice agent data, from 8 kHz phone calls to 48 kHz studio audio.
- [Speech Data RFP Template: What to Ask Vendors Before You Buy](https://www.speechdata.ai/guides/speech-data-rfp-template): A copy-paste speech data RFP template: what to ask vendors about spec, consent, license and price per hour, plus a scoring rubric and red flags to watch.
- [Medical Speech Recognition: What Training Data It Needs](https://www.speechdata.ai/guides/medical-speech-recognition-data): Why medical speech recognition misses drug names, doses and exam-room talk, what data each use case needs, and where to source it.
- [Full-Duplex Speech Data for Voice Agents and Speech-to-Speech AI](https://www.speechdata.ai/guides/full-duplex-conversational-speech-data): What full-duplex and speech-to-speech models need from training data, what Moshi and dGSLM trained on, and a checklist for buying dual-channel audio.
- [ASR Test Sets: How to Measure Real Speech Recognition Accuracy](https://www.speechdata.ai/guides/asr-test-set): Why ASR benchmarks overstate production accuracy, and how to build a held-out test set: what audio to include, how many speakers, splits and slicing.
- [Speech Data for Low-Resource Languages: Why It's Scarce and How to Get It](https://www.speechdata.ai/guides/speech-data-for-low-resource-languages): Why speech data for low-resource languages is scarce, why open corpora fall short, and how commercial collection closes the gap for voice AI teams.
- [How to Buy Speech Data: The Complete Buyer's Guide](https://www.speechdata.ai/guides/how-to-buy-speech-data): How to buy speech data: scope requirements, audit vendors in 30 minutes, benchmark pricing, and negotiate the contract. A practical guide for ML teams.
- [Speech Data Collection Companies: Who Records Custom Datasets](https://www.speechdata.ai/blog/speech-data-collection-companies): Speech data collection companies compared: who records speech to order, how (remote, studio or on-site), which languages they claim, and what to ask.
- [Spanish and Portuguese Speech Data: Which Variety Do You Need?](https://www.speechdata.ai/blog/spanish-portuguese-speech-data): Spain or Mexico, Portugal or Brazil: how Spanish and Portuguese varieties differ for speech models, which to buy for each market and how to test them.
- [Off-the-Shelf vs Custom Speech Data: How to Choose](https://www.speechdata.ai/blog/off-the-shelf-vs-custom-speech-data): When to license off-the-shelf speech data, when to commission custom collection, and how to combine both, with a decision table and worked costs.
- [Arabic Speech Recognition: Dialects, Data and Spelling](https://www.speechdata.ai/blog/arabic-speech-recognition-dialects): Why Arabic ASR struggles with dialect speech, what the MGB challenges measured, how to spell dialect transcripts and which dialect data each market needs.
- [Switchboard Corpus, Fisher and CALLHOME: Commercial Alternatives](https://www.speechdata.ai/blog/switchboard-fisher-alternatives): What the Switchboard corpus, Fisher and CALLHOME contain, how LDC licenses them, and two-channel conversational speech you can use commercially.
- [Dual-Channel Audio for Call Analytics and Voice AI](https://www.speechdata.ai/blog/dual-channel-audio-speech-ai): Dual-channel audio puts each speaker of a call on its own channel. What that enables for call analytics and voice AI, how to record it, and what to check.
- [Deepfake Audio Detection Data: Why Real Speech Matters](https://www.speechdata.ai/blog/deepfake-detection-training-data): Audio deepfake detection fails when the real speech in training doesn't match production. Public datasets, their licenses and a spec for the bona fide set.
- [Appen Alternatives for Speech Data](https://www.speechdata.ai/blog/appen-alternatives-speech-data): Appen alternatives for ASR and TTS training data: what Defined.ai, LXT, Shaip, FutureBeeAI, Nexdata, DataForce, TELUS Digital and SpeechData.ai offer.
- [Speech Data Providers Compared: How to Choose One in 2026](https://www.speechdata.ai/blog/speech-data-providers): Speech and voice data providers compared: open datasets, commercial vendors such as Appen, Defined.ai, LXT and Shaip, and a checklist for choosing one.
- [How to Buy AI Training Data: A Practical Playbook (2026)](https://www.speechdata.ai/blog/how-to-buy-ai-training-data): How to buy AI training data without getting burned: provider types, pricing models, consent and license checks, red flags, and a procurement checklist.

## Contributors

- [Remote speech projects for freelancers](https://www.speechdata.ai/contributors): register in the Spirelight portal for project invitations; voice recording, transcription and speech annotation; registration does not guarantee work
- [Voice recording preparation](https://www.speechdata.ai/contributors/voice-recording): practical preparation for a remote speech recording task
- [Transcription practice](https://www.speechdata.ai/contributors/transcription): a practice exercise and guideline checklist for new speech contributors

## Optional

- [Verbatim vs. Clean Verbatim Transcription: What Speech AI Needs](https://www.speechdata.ai/blog/verbatim-vs-clean-verbatim-transcription): Verbatim, clean verbatim and intelligent verbatim transcription compared for ASR and TTS data, with a worked WER example and a guideline checklist.
- [Inter-Annotator Agreement in Speech Data: Kappa, Alpha and WER](https://www.speechdata.ai/blog/inter-annotator-agreement-speech): How to measure inter-annotator agreement on speech data: Cohen's and Fleiss' kappa, Krippendorff's alpha, transcript WER and DER, with a worked example.
- [Forced Alignment: How to Get Word Timestamps for Speech Data](https://www.speechdata.ai/blog/forced-alignment-speech-data): How forced alignment turns a transcript into word and phone timestamps: MFA, WhisperX and MMS compared, why alignment fails, and what to specify.
- [How to Choose an Audio Annotation Vendor](https://www.speechdata.ai/blog/audio-annotation-services): A buyer's guide to choosing an audio annotation vendor: how annotation is priced, how to judge quality, what to ask, and when to build in-house.
- [The Data Behind Voice Cloning: Recordings, Consent, and Quality](https://www.speechdata.ai/blog/voice-cloning-training-data): Voice cloning training data: recording quality, speaker consent and licensing checks for teams evaluating data before building a voice model.
- [Wake Word Detection: How It Works and How to Train a Custom Wake Word](https://www.speechdata.ai/blog/wake-word-detection): Wake word detection explained: how always-on keyword spotting works, false accepts vs false rejects, tools like Porcupine and openWakeWord, and training data.
- [Synthetic Speech Data vs Real Recordings: What Trains Better Models](https://www.speechdata.ai/blog/synthetic-vs-real-speech-data): When synthetic speech data helps, when it hurts, and how to combine it with real recordings. A practical comparison for teams training ASR and voice AI.
- [Speech Data Collection: The Complete 2026 Guide](https://www.speechdata.ai/blog/speech-data-collection-guide): Speech data collection from spec to delivery: recruitment, consent and GDPR, recording setups, QA, transcription, costs, and the build-vs-buy decision.
- [How to Fine-Tune Whisper: Data, Setup and Evaluation](https://www.speechdata.ai/blog/how-to-fine-tune-whisper): Prepare a Whisper fine-tune: choose a baseline, check audio-text pairs, separate evaluation and configure checkpoint selection before scaling a run.
- [Speech Data for Voice Agents and Call-Center AI](https://www.speechdata.ai/blog/speech-data-for-voice-agents): Speech data for voice agents: two-party phone audio, domain vocabulary and test sets, sourced from licensed datasets or role-played calls.
- [Voice Data Privacy: Consent, PII, and GDPR for Speech Datasets](https://www.speechdata.ai/blog/voice-data-privacy-compliance): How to handle consent, personal data, and PII redaction in speech datasets. A practical compliance primer for teams collecting or buying voice data.
- [Accent and Dialect Diversity in Speech Data](https://www.speechdata.ai/blog/accent-and-dialect-speech-data): Plan accent and dialect coverage for ASR with an evaluation checklist, a hypothetical recording brief, and limits on what new data can fix.
- [Word Error Rate (WER) Explained: How to Measure Speech Recognition Accuracy](https://www.speechdata.ai/blog/word-error-rate-explained): Word error rate (WER) explained: the (S+D+I)/N formula with a worked example, computing WER with jiwer, normalization pitfalls, CER, and typical WER bands.
- [The Best ASR Models in 2026: Open-Source and Commercial Compared](https://www.speechdata.ai/blog/best-asr-models-2026): The best ASR models in 2026 compared: Whisper, NVIDIA Canary and Parakeet, Deepgram, AssemblyAI. Accuracy, languages, licenses, and how to choose.
- [Speaker Diarization: How 'Who Spoke When' Actually Works](https://www.speechdata.ai/blog/speaker-diarization-guide): Speaker diarization explained: the VAD → embedding → clustering pipeline, end-to-end models, tools like pyannote and WhisperX, and the DER metric.
- [15 Best Open-Source Speech Datasets for ASR and TTS (2026)](https://www.speechdata.ai/blog/best-open-source-speech-datasets): The 15 best open-source speech datasets for ASR and TTS in 2026: sizes, languages, licenses, and honest notes on where open data falls short.
- [Code-Switching: What It Is and Why It Breaks Speech Recognition](https://www.speechdata.ai/blog/code-switching-speech-recognition): What is code switching? Definition, linguistic types, real examples like Hinglish and Spanglish, and why it remains ASR's hardest everyday problem.
- [What Is Voice Activity Detection (VAD)? How It Works and Why It Matters](https://www.speechdata.ai/blog/what-is-voice-activity-detection): Voice activity detection (VAD) explained: how energy-based and ML-based VAD works, where it sits in ASR and voice-agent pipelines, and how to evaluate it.
- [What Is Speech Synthesis? From Formant Synthesis to Neural TTS](https://www.speechdata.ai/blog/what-is-speech-synthesis): What is speech synthesis? How TTS evolved from formant and concatenative systems to neural models like VITS, and why training data now decides quality.
- [What Is Audio Annotation? Types, Process, and Best Practices](https://www.speechdata.ai/blog/what-is-audio-annotation): Audio annotation explained: transcription, diarization, timestamps, and acoustic events, plus QA methods, tools, cost drivers, and build-vs-buy tradeoffs.
- [Conversational vs. Scripted Speech Data: Which Does Your Model Need?](https://www.speechdata.ai/blog/conversational-vs-scripted-speech-data): Conversational speech datasets vs. scripted read speech: acoustic differences, the WER gap on spontaneous speech, and which to use for ASR, TTS, and voice AI.
- [What Is Automatic Speech Recognition (ASR)? A Practical Guide](https://www.speechdata.ai/blog/what-is-automatic-speech-recognition): What is automatic speech recognition? A practical ASR definition, how modern speech recognition works, WER explained, and the training data behind it.
- [How Much Audio Data Do You Need to Train an ASR Model?](https://www.speechdata.ai/blog/how-much-audio-data-to-train-asr): Size an ASR data project with a held-out test set, a staged pilot and clear acceptance criteria. Includes published examples and a collection brief.
- [TTS Datasets: How to Choose Training Data for Text-to-Speech Models](https://www.speechdata.ai/blog/text-to-speech-datasets): What makes a good TTS dataset: audio quality, hours per voice, phonetic coverage, prosody, and licensing, from VITS fine-tuning to zero-shot voice cloning.
- [About](https://www.speechdata.ai/about): who runs SpeechData.ai and how the data is collected