LIVE
Agentic test processes, LLM benchmarks, and other notes on agentic codingCFTC issues second warning to prediction markets on cookie-cutter self-certificationsDesign Is CompromiseShow HN: I mapped every US golf courseRussia’s Sberbank to launch crypto trading infrastructure this yearAn Inside Look at the Token Reseller MarketInflect-Micro-v2: complete voice in 9.36M parametersMira Murati’s Inkling AI Model Review: Best Open-Source Model in the WestPark by Robot at London Gatwick AirportDeepSeek pause fundraise after comments on compute gap to US leaked (transcript) [pdf]Bitcoin OG selling eases as dormant BTC movement hits 4-year low: GalaxyGoogle Discloses $94.1B in SpaceX Stock, Marking 6% StakeShow HN: Brolly, a plain-text weather forecast siteHere’s what happened in crypto todayGo Analysis Framework: modular static analysis by go teamThe AI jobs apocalypse probably isn't coming anytime soonSouth Korea’s largest bank to launch payment service on JPMorgan’s KinexysThird Drone Shot Down in Three Days in Romanian TerritoryOpenAI did not notice Hugging Face hack for a weekBitMart to wind down exchange, end trading by Aug. 26Htmx 4.0, the first JavaScript library to release exclusively on the Game BoyTurn And Face The Strange: Fly.io is betting on computers for AI agentsBinance ‘red teams’ its own staff every month to keep hackers outKill The Cookie BannerNvidia, Microsoft, Meta warn against overregulating open-weight modelsRobinhood in talks with Crypto.com over prediction markets: WSJCarrie is just trying to make a friend in the new trailerShow HN: Audio Player with "Binaural Beats" tuned to the same key as your musicBitcoin advocacy group to join US State Department’s ‘digital freedom’ programMarvel announces Ryan Gosling as Ghost Rider, a new Black Panther, and more
Awesome Free Compute
Home / Media

🎙️ The Free Audio Generation Guide (2026)

Music · voice (TTS + cloning) · transcription — the $0 stack for a solo creator who monetizes, with a hard focus on the thing every other list glosses over: commercial licensing.

Verified June 2026. Numbers and licenses change fast. Claims are marked ✅ verified (read from a primary source — pricing page, license file, or model card) or ⚠️ verify (page was JS/login-gated or the free-plan detail wasn't exposed). Researched with a multi-model sweep (GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.8) and reconciled against official sources.

This is a companion to the main Awesome Free Compute list and the Free Token-Maxxing Guide (LLMs/tokens). This file is the audio half: anything you'd need to score, narrate, and caption a video for $0. Once you've generated the audio, the Media Post-Production Guide 🎬 covers finishing it — stem separation, mastering, loudness, dubbing.


Contents


TL;DR

The $0 audio stack for a creator who needs commercial rights:

Need Use Why
🎵 Free + commercial full songs (with vocals) ACE-Step (Apache-2.0) on free GPU The only open full-song model with vocals and a clean commercial license; runs on ~8 GB.
🎹 Free commercial instrumental, zero setup Adobe Firefly Generate Soundtrack (2 lifetime free) or Stable Audio Open (<$1M rev) Hosted, royalty-free, WAV, YouTube-monetizable — but Firefly free = only 2 generations.
🗣️ Free TTS narration (incl. Hindi/Hinglish) Azure Speech F0 (0.5M chars/mo) or Google Cloud TTS (1M chars/mo) Perpetual free tier, full commercial rights, excellent hi-IN/en-IN voices.
🧬 Free + commercial voice cloning GPT-SoVITS / F5-TTS / CosyVoice (MIT/Apache) Clone your own voice once, reuse forever, monetize legally. (Avoid XTTS-v2 + Fish Speech.)
📝 Free transcription + karaoke timestamps WhisperX (self-host) or Groq Whisper (hosted) WhisperX = word-level alignment + diarization for lyric/subtitle timing.
🇮🇳 Best free Hindi/Hinglish STT Azure continuous LID [hi-IN,en-IN] or Deepgram Nova-3 multi The two options that actually document Hindi code-switching.
♾️ Unlimited + offline (Mac) whisper.cpp / MLX-Whisper / faster-whisper + Kokoro / Piper TTS MIT, on-device on Apple Silicon, no quota, private.

One-liner: Song → ACE-Step (or Firefly for an instrumental bed). Voice → Azure F0 / Kokoro, clone your own voice with GPT-SoVITS. Captions & lyric timing → WhisperX. Hinglish → Azure LID or Deepgram Nova-3. All $0, all commercial-clean.


The licensing landmines

For a monetized channel, "free to use" almost never means "free to monetize." These are the traps that look free but are not commercially usable on their free/open terms:

⚠️ Looks free, isn't commercial Why What it blocks
Suno free/Basic plan ✅ Suno owns the song; personal/non-commercial only; not retroactive when you upgrade Monetized YouTube
Meta MusicGen / AudioGen / MAGNeT ✅ weights are CC-BY-NC 4.0 Any commercial output
Coqui XTTS-v2 CPML is a non-commercial model license Monetized narration/cloning
Fish Speech / OpenAudio ✅ research license — "commercial use requires a separate license" Monetized narration/cloning
ElevenLabs free tier & Scribe free ✅ commercial license only from Starter ($5/mo) Monetized VO + transcripts
Soundful free/Standard, Loudly free ✅ commercial rights are paid-plan only Monetized music
NVIDIA Parakeet / Canary (STT) License is fine (CC-BY-4.0) but no Hindi Hindi/Hinglish work

Green-light — free and commercial: ACE-Step (Apache-2.0), YuE (Apache-2.0), Stable Audio Open / Open Small (under $1M revenue), Adobe Firefly Generate Soundtrack, Azure Speech F0, Google Cloud TTS, Kokoro / Piper / F5-TTS / GPT-SoVITS / CosyVoice (MIT/Apache), and the whole Whisper family (MIT) + AI4Bharat IndicConformer (MIT).


1. Music generation

1.1 Open weights — self-host = truly free + commercial

The cleanest commercial path. Run them on a free Colab T4, Kaggle, HF ZeroGPU, or a beefy local GPU (see §4).

Model License Vocals? Max length VRAM Free GPU?
ACE-Step v1 (3.5B) 🥇 Apache-2.0 Yes ~4 min coherent ~8 GB with --cpu_offload ✅ Colab T4 / Kaggle / HF Space
YuE Apache-2.0 ✅ (attribution encouraged) Yes, full songs several min (30s segments) 24 GB (≤2 sessions); 80 GB for full ⚠️ marginal on T4; ZeroGPU/quantized UI
Stable Audio Open 1.0 Stability Community ✅ (<$1M rev, register) No realistic vocals 47 s 16–48 GB ideal ⚠️ short runs on T4; ZeroGPU safer
Stable Audio Open Small Stability Community ✅ (<$1M rev) No vocals 11 s low ✅ T4 / ZeroGPU
Meta MusicGen / MAGNeT ⚠️ CC-BY-NC 4.0 No vocals 10–30 s small models fit T4 ✅ but non-commercial

Pick: ACE-Step is the standout — Apache-2.0, generates a full song with vocals from lyrics + a style prompt, duration control, lyric editing/remix, and a memory-optimized path down to ~8 GB. Ready-made free Space: huggingface.co/spaces/ACE-Step/ACE-Step. Use your own lyrics, keep the project files, and run a similarity check before publishing.

1.2 Hosted apps with free tiers

App Free allowance Vocals? Commercial on free? Notes
Adobe Firefly — Generate Soundtrack 🥇 2 lifetime free premium uses ✅ No (instrumental) ✅ Yes — royalty-free, YouTube/TikTok OK 5 s–5 min, WAV export, Content Credentials attached. Cleanest hosted $0 commercial path, but only 2 generations.
Suno free/Basic exists Yes ❌ No — Suno owns; non-commercial Make the song on a paid plan if you want to monetize it.
Soundful Standard/free Mostly instrumental ❌ No Standard tracks explicitly not for commercial use.
Loudly free tier Yes ❌ No (paid-only rights) License grants monetization under paid plans.
Stable Audio (hosted) free plan, monthly Instrumental/SFX ⚠️ unverified for free Use Stable Audio Open self-host for clear rights.
Mubert / SOUNDRAW / Beatoven varies Instrumental/loops ⚠️ unverified for free SOUNDRAW's "All Plans" license is promising but its free-plan schedule wasn't exposed; verify in-app at download.
Google MusicFX free experimental n/a ⚠️ unclear No published commercial-output license; not recommended for monetized music.

1.3 Chinese music apps — caveat

Mureka (Kunlun SkyMusic), Tiangong SkyMusic, Haimian Music (ByteDance), Tencent Lingyin/TME, NetEase Tianyin, ACE Studio. Marketing often claims "full commercial rights," but the free-plan ToS were JS/PDF/login-gated and could not be verified this pass, and several need a +86 phone number. For a monetizing creator, prefer the open ACE-Step model — its Apache-2.0 license is unambiguous, versus trusting an in-app license you can only read at generation time.

Verdict (music):ACE-Step (open, Apache-2.0, vocals, free GPU) → ② YuE if you need richer vocal songs and can find a 24 GB+ free GPU → ③ Adobe Firefly Generate Soundtrack for a quick commercial-safe instrumental bed. Avoid for monetized YouTube: Suno free, Soundful/Loudly free, MusicGen.


2. Text-to-speech and voice cloning

2.1 Hosted free tiers

Engine Free quota Hindi/Hinglish Commercial on free? Notes
Azure Speech F0 🥇 0.5M chars/mo neural ✅, perpetual Excellent (hi-IN Swara/Madhur, en-IN) ✅ Yes, no watermark Best $0 narration; you already use it.
Google Cloud TTS 1M chars/mo ✅ (Standard/WaveNet/Neural2) Strong hi-IN ✅ Yes, no attribution Highest perpetual free quota.
Amazon Polly 1M neural chars/mo, first 12 mo only Good ✅ Yes (then expires) Free-tier clock runs out after a year.
ElevenLabs 10k chars/mo Excellent ❌ No (+ attribution) Commercial unlocks at Starter $5/mo.
PlayHT / Murf / Speechify trial / very limited Good ❌ No Free tiers prohibit commercial use.
Deepgram Aura / Cartesia / Resemble sign-up credit English-led ✅ until credit ends No perpetual commercial free tier.

2.2 Open weights — free + commercial

Run on a free Colab T4, HF ZeroGPU, or locally on an Apple Silicon Mac (many run on CPU).

Model License Cloning? Hindi/multiling Runs on
Kokoro-82M 🥇 Apache-2.0 No (uses voice embeddings) Hindi added in v1.0 CPU + any GPU; tiny
Piper MIT No (fine-tune) native Hindi voices CPU-optimized, edge
F5-TTS 🥇 MIT Yes (15s zero-shot) cross-lingual Mac / Colab T4
GPT-SoVITS 🥇 MIT Yes (1-min few-shot, best) cross-lingual (EN→Hindi) Apple Silicon / T4
CosyVoice / CosyVoice2 (Alibaba) Apache-2.0 Yes (3s zero-shot) cross-lingual CPU/GPU
MeloTTS MIT No no Hindi (EN/ZH/ES/FR/KR/JP) CPU fast
Suno Bark MIT audio-prompt only multilingual needs GPU
Coqui XTTS-v2 ⚠️ CPML non-commercial Yes (6s) native Hindi don't monetize
Fish Speech ⚠️ research license Yes (10s) yes don't monetize

2.3 Hindi/Hinglish + cloning legality

  • Best Hindi/Hinglish narration: hosted Azure F0 or Google Cloud TTS — they handle code-switched, transliterated text far better than most open models.
  • Best $0 local: Kokoro (narration) and GPT-SoVITS (clone your own voice once, then unlimited Hindi/Hinglish output, MIT, monetizable).
  • Cloning legality: open models won't stop you cloning anyone — but monetizing someone else's voice violates rights of publicity/likeness. Only clone your own voice (or one you have written permission for). YouTube also requires disclosure of realistic AI-generated content.

Verdict (voice): narration → Azure F0 / Google TTS (hosted) or Kokoro (local); cloning → GPT-SoVITS (best, MIT) or F5-TTS / CosyVoice for short-sample zero-shot. Avoid XTTS-v2 and Fish Speech for monetized work.


3. Transcription and subtitles

3.1 Open weights — self-host on free GPU/Mac

Engine License Hindi/code-switch Timestamps / diarization Runs on
Whisper large-v3 / turbo MIT Hindi ✅; Hinglish ~OK segment native Colab/Kaggle; Mac
faster-whisper (CTranslate2) MIT = Whisper word ✅ CPU int8 on Mac; T4 (~2.9 GB)
WhisperX 🥇 BSD-2 = Whisper + Hindi wav2vec2 align word ✅ + diarization ✅ Mac CPU int8; T4
whisper.cpp MIT = Whisper word ✅ Apple Silicon (Metal/CoreML)
MLX-Whisper MIT = Whisper word ✅ Apple Silicon (MLX)
AI4Bharat IndicConformer 600M MIT Hindi + 22 Indian langs (Devanagari) CTC timestamps GPU/CPU
NVIDIA Parakeet / Canary CC-BY-4.0 ✅ ⚠️ no Hindi (EN/European) word ✅ GPU (NeMo/CUDA)
distil-whisper MIT ⚠️ English-only word ✅ any

Key finding: the models at the top of the English ASR leaderboard — NVIDIA Parakeet/Canary, IBM Granite, Cohere — do not support Hindi. For Hinglish, Whisper large-v3 (via faster-whisper/WhisperX) is the best free open engine.

3.2 Hosted free tiers / credits

Service Free Hindi/code-switch Diarization Commercial?
Groq Whisper-v3/turbo free tier, 25 MB file cap, fast Hindi ✅ (Whisper) no
Deepgram Nova-3 🥇 $200 credit, no card hi + multi code-switch
AssemblyAI $50 credit, no card Universal-2 Hindi (10–25% WER) ✅ (+$0.02/hr)
Azure Speech F0 ~5 audio-hr/mo ⚠️ continuous LID [hi-IN,en-IN] ✅ (batch)
Google Cloud STT $300/90d + ~60 min/mo ⚠️ hi-IN + Indian langs
ElevenLabs Scribe 10k credits/mo 90+ langs ✅ (≤32 spk) ❌ free = non-commercial

3.3 Hindi / Hinglish — the hard case

Casual Hindi-English code-switch is the toughest case (plain local Whisper struggles). Ranked free options:

  1. Azure continuous LID [hi-IN,en-IN] — verified-good, per-segment language switching, ~5 free hrs/mo. (Your existing baseline.)
  2. Deepgram Nova-3 multi — the only hosted free-credit option that explicitly documents Hindi plus code-switching ($200 credit).
  3. Self-host Whisper large-v3 (faster-whisper/WhisperX) for unlimited free Hinglish; AI4Bharat IndicConformer or IIT-Madras whisper-hindi for clean Devanagari Hindi (these are weaker on English mixing).

3.4 Subtitles & karaoke timing

WhisperX is the best free tool: word-level forced alignment + speaker diarization + subtitle export, on Colab or Mac. For Hindi songs, pair it with a Hindi wav2vec2 alignment model. Mac-local alternatives with word timings: faster-whisper (word_timestamps=True), MLX-Whisper, whisper.cpp.

Verdict (STT): general → self-host Whisper large-v3 (or Groq for zero setup); Hindi/Hinglish → Azure LID / Deepgram Nova-3; karaoke timing → WhisperX.


4. Free GPU to run it all

The open music/voice/STT models above are commercial-clean only if you self-host — and you can do that for $0:

  • HF Spaces / ZeroGPU — using existing ZeroGPU Spaces (ACE-Step, MusicGen, Whisper demos) is free; authenticated free accounts get ~5 min GPU/day on a 48 GB Blackwell slice (PRO = 40 min). Great for one-off generations.
  • Google Colab (free) — ~16 GB T4. Comfortable for ACE-Step (offload), Stable Audio Open Small, MusicGen-small, and faster-whisper/WhisperX large-v3.
  • Kaggle Notebooks (free) — weekly GPU quota (T4×2 / P100); best for batch transcription or longer renders. Verify the current quota in your account.
  • Mac-local (Apple Silicon) — fully offline, unlimited minutes, private: whisper.cpp / MLX-Whisper / faster-whisper (STT) and Kokoro / Piper (TTS) all run well on an M-series CPU/GPU.

For more on these venues (and serverless GPU like Modal), see the main Awesome Free Compute list.


5. The $0 creator pipeline

End-to-end for a music-video creator, every step commercial-clean and free:

  1. SongACE-Step (Apache-2.0) on a free GPU for a full track with vocals from your own lyrics; or Adobe Firefly Generate Soundtrack for a quick instrumental bed.
  2. Narration / VOAzure Speech F0 Hindi/Indian-English voices (0.5M chars/mo), or Kokoro locally for unlimited.
  3. Your signature voice — clone your own voice once with GPT-SoVITS (MIT), then reuse it forever for hooks/VO.
  4. Lyric timing / karaoke captionsWhisperX word-level timestamps on the vocal stem.
  5. Hinglish transcription / subtitlesAzure continuous LID [hi-IN,en-IN] (your baseline) or Deepgram Nova-3 multi ($200 credit).

Total cost: $0. Total commercial exposure: none, as long as you stick to the green-light tools in the licensing landmines.


Sources

Primary sources fetched/verified June 2026 (representative):

  • Music — open weights: github.com/ace-step/ACE-Step + huggingface.co/ACE-Step/ACE-Step-v1-3.5B (Apache-2.0); github.com/multimodal-art-projection/YuE (Apache-2.0); huggingface.co/stabilityai/stable-audio-open-1.0 + /stable-audio-open-small + stability.ai/license; huggingface.co/facebook/musicgen-small (CC-BY-NC).
  • Music — hosted: adobe.com/products/firefly/features/ai-music-generator.html (2 lifetime, commercial); help.suno.com/en/articles/2746945 + suno.com/terms (free = non-commercial); soundful.com/pricing + /license; loudly.com/license-agreement.
  • TTS — hosted: azure.microsoft.com/pricing/details/speech (F0 0.5M chars/mo); cloud.google.com/text-to-speech/pricing (1M chars/mo); elevenlabs.io/pricing (free non-commercial); aws.amazon.com/polly/pricing.
  • TTS — open weights: huggingface.co/hexgrad/Kokoro-82M (Apache-2.0); github.com/SWivid/F5-TTS (MIT); github.com/RVC-Boss/GPT-SoVITS (MIT); github.com/FunAudioLLM/CosyVoice (Apache-2.0); coqui/XTTS-v2 (CPML, non-commercial); github.com/fishaudio/fish-speech (research license); Piper / MeloTTS / Bark (MIT).
  • STT — open weights: github.com/openai/whisper (MIT); github.com/SYSTRAN/faster-whisper; github.com/m-bain/whisperX; github.com/ggml-org/whisper.cpp; huggingface.co/nvidia/parakeet-tdt-0.6b-v2 + /canary-1b-flash (CC-BY-4.0, no Hindi); huggingface.co/ai4bharat/indic-conformer-600m-multilingual (MIT); huggingface.co/spaces/hf-audio/open_asr_leaderboard.
  • STT — hosted: console.groq.com/docs/speech-to-text; deepgram.com/pricing + developers.deepgram.com (Nova-3 multi Hindi); assemblyai.com/pricing; learn.microsoft.com/azure/ai-services/speech-service/language-identification (continuous LID); cloud.google.com/speech-to-text/pricing; elevenlabs.io/pricing/api (Scribe).
  • Free GPU: huggingface.co/docs/hub/spaces-zerogpu; research.google.com/colaboratory/faq.html; kaggle.com/docs/notebooks.

Quotas/licenses are point-in-time (June 2026) and change frequently — verify the free-plan terms at signup/download before relying on them, especially for commercial use. Corrections via PR welcome.


📝 Spotted a stale quota or a license that changed? This guide is open source — edit it on GitHub.